Google has announced that its newest chips for training AI models are more powerful and efficient than competing chips from Nvidia, and have been used to train its large language models in a supercomputer.

Its 4th generation tensor processing unit (TPU) is better equipped for AI training than its predecessors, and offers 1.7 times speed improvements over Nvidia’s competing GPU the A100, the company claimed.

Google outlined the technical details of TPU v4 in a research paper published on Tuesday. It stated that the chip formed the basis for its third supercomputer specifically intended for machine learning (ML) models, which contains 4,096 TPUs for ten times greater power overall than previous clusters.

High-performance chips such as the TPU are necessary components for supercomputers, which are seeing particularly high use right now as the hardware used to train large language models (LLMs) for generative AI.

Google trained its own 540 billion-parameter PaLM LLM using two TPU v4 supercomputer clusters, which partially forms the basis for Bard, its answer to ChatGPT.

Last month, Google also announced that the AI image generation program Midjourney was trained using Google TPUs, GPU VMs, and Google Cloud infrastructure.

The firm highlighted the TPU v4's optical circuit switches, a high-speed, low-power alternative to Infiniband node connectors that can be used to dynamically reconfigure the chip's topology to fit user needs.

The TPU also uses 1.3-1.9 times less power than Nvidia’s A100 chips, a crucial statistic in the interest of keeping energy costs low and maintaining carbon budgets for energy centres and supercomputers.

Amid ongoing energy price volatility, Gartner predicted that data centres could face cost increases of up to 40%, and the UK government reportedly held crisis talks with data centre operators in October over blackout concerns.

Google TPUs are used throughout Google Cloud data centres, and the firm stated that the 4th generation chips in Google Cloud warehouse computers use three times less energy, producing 20 times fewer CO2 emissions than contemporary data centres.

Gartner additionally identified sustainability as the key strategic technology trend for 2023, and with this in mind, the TPU v4 could make Google Cloud an attractive proposition for firms looking to train AI with minimal carbon impact.

Absent from the paper was any mention of Nvidia’s H100, the firm’s new AI accelerator of choice announced in March 2022.

Nvidia has stated that the H100 delivers 6.7 times greater performance than its predecessors in ML benchmarks, with the A100 delivering 2.5 times performance improvements by the same measurements.

Big tech has flocked to Nvidia for AI hardware, and the GPU giant has even announced that its DGX Cloud supercomputer service will soon be available through Google Cloud.

It remains one of the most trusted choices for powerful and energy-efficient hardware for AI, and provided the chips for Microsoft’s supercomputer used to train OpenAI’s GPT-4 model.

An Nvidia spokesperson told ITPro that the firm's H100 chip delivers the best performance across categories including natural language understanding, recommender systems, computer vision, speech AI and medical imaging. 

“Three years ago when we introduced A100, the AI world was dominated by computer vision. Generative AI has arrived,” said NVIDIA founder and CEO Jensen Huang in a blog post.

“This is exactly why we built Hopper, specifically optimized for GPT with the Transformer Engine. Today’s MLPerf 3.0 highlights Hopper delivering 4x more performance than A100."

This article was updated to include a statement from Nvidia. 

The decision by Italian regulators to ban access to ChatGPT has drawn criticism from industry experts, branding it a hindrance to innovation and a distraction from legitimate concerns around AI and privacy.

Italian authorities banned access to ChatGPT at the end of March. The degree to which the technology collects data for algorithmic training purposes was deemed to have no legal basis.

The lack of age verification embedded in the technology, which could shield underage users from inappropriate content, was also highlighted as a key issue.

The Italian data protection authority, Garante per la protezione dei dati personali, issued a decision on 31 March declaring that it would temporarily suspend the processing of Italian user data by OpenAI.

OpenAI has restricted access to ChatGPT in Italy while it works with Garante to establish an understanding.

Those who purchased ChatGPT Plus subscriptions in March will receive a refund, and all subscribers in the region have had their recurring payments paused.

The ban prompted debate among AI and legal experts who have questioned the legitimacy and wiseness of the decision.

Andy Patel, researcher at WithSecure, called the ruling an “overreaction” and warned that it could put Italy at a disadvantage when it comes to AI development.

“ChatGPT is a useful tool that enables creativity and productivity - by shutting it off, Italy has cut off perhaps the most important tool available to our generation," he said.

"All companies have security concerns, and, of course, employees should be instructed to not provide ChatGPT and similar systems with company-sensitive data. Such policies should be controlled by individual organisations and not by the host country.”

Others have suggested that the ban will not last, and is a distraction from the wider issues of privacy and security that must be addressed when it comes to large language models (LLMs).

“‘Banning’ these models - whatever that term means in this context - is simply encouraging more perfidy on the part of these companies to restrict access and concentrates more power in the hands of tech giants who are able to sink the money into training such models,” said Erick Galinkin, principal artificial intelligence researcher at Rapid7.

“Rather, we should be looking for more openness around what data is collected, how it is collected, and how the models are trained.”

Concerns around the centralisation of generative AI in the hands of big tech at present have led to calls for a 'democratisation' of the technology.

AI model leakers have called for the stolen Meta LLM LLaMA to be stored on Bitcoin to maintain free distribution, while AWS and Hugging Face partnered to improve access to models.

ChatGPT raised eyebrows last month when a privacy flaw exposed users' chatbot interactions, the first real blow to the company's image since its early GPT-3 tests were branded 'too dangerous' for public use.

Michael Covington, VP of strategy at software firm Jamf, noted that there is value in a pause to ensure that AI technology proceeds in a controlled manner.

“That said, I get concerned when I see attempts to regulate common sense and force one 'truth' over another,” he added.

“At Jamf, we believe in educating users about data privacy, and empowering them with more control and decision-making authority over what data they are willing to share with third parties. Restricting the technology out of fear for users giving too much to any AI service could stunt the growth of tools like ChatGPT, which has incredible potential to transform the ways we work.”

The Italian authority also asked for OpenAI to provide notice of measures implemented to comply with its order, or face a fine as large as €20 million ($21 million) or 4% of the company’s worldwide annual turnover.

Could other countries ban ChatGPT?

In becoming the first European country to bar access to ChatGPT, Italy has set a precedent that other countries could follow.

While the ban is in place it has joined the likes of Russia, Iran, and North Korea which all ban ChatGPT as part of wider internet censorship.

OpenAI has its services geoblocked in China, meaning that businesses and consumers are unable to access ChatGPT and DALL·E in the region.

Domestic companies such as Baidu have unveiled alternative chatbots, but as yet none have demonstrated abilities on par with OpenAI’s GPT-4.

Reuters reported that the Irish Data Protection Commission (DPC) and French Commission nationale de l'informatique et des libertés (CNIL) are in discussions with their Italian counterparts to establish the basis for the decision.

If the discussion proves convincing, regulatory bodies across the EU could soon require further privacy commitments from OpenAI.

“The Garante’s decision is a timely reminder that the excitement around generative AI must be tempered with caution,” Will Richmond-Coggan, a partner at national law firm Freeths specialising in privacy and technology, told IT Pro.

“OpenAI now have an opportunity to show how it has been gathering and using personal data to train its tool in a way that is compatible with GDPR. If it can’t, it seems likely that other European supervisory authorities may follow Italy’s lead.”

Richmond-Coggan noted that the focus on the impact of ChatGPT on children and questions around age-appropriate content will keep this regulatory interest alive, and that this is a issue with which AI developers will have to contend for some time.

“With the prospect of future legislation in the UK and Europe directed both to online harms and more focused on regulating AI technologies, this is likely to be only the first of a large number of regulatory hurdles,” he added.

The government’s recently released AI whitepaper identifies fairness as an attribute necessary for AI innovation, and specifies that this pertains to the compliance of AI systems with laws such as UK GDPR.

The government has also sought to outline the transparency and redress requirements by which firms operating in the space will have to abide.

Currently, AI models operate largely on a ‘black box’ principle, in which users or even business partners have little to no oversight of the data used to train models, or what data is processed to improve models.

But the government’s approach has also been explicitly pro-innovation, and was praised by industry leaders synch as Microsoft UK CEO Clare Barclay as a “commitment to being at the forefront of progress”.

In the near future, AI companies could instead be compelled to shed more light on their data scraping, processing, and training activities.

A recent panel of experts noted that greater AI transparency is necessary to avoid future regulatory penalities, and this debate is will only intensify in the coming years.

Intel has laid out a roadmap for establishing product leadership in the processor market by 2025, alongside a goal of democratising AI under a consolidated range of AI-optimised hardware and software.

Core to its proposition is a diverse range of products including central processing units (CPUs), graphics processing units (GPUs), and dedicated AI architecture alongside open-source software improvements.

Businesses can expect to benefit from fourth-generation ‘Sapphire Rapids’ Xeon CPUs immediately, with the fifth-generation Xeon codenamed ‘Emerald Rapids’ set for a Q4 2023 release. This will be followed in 2024 by two processors known as Granite Rapids and Sierra Forest.

Sapphire Rapids can deliver up to ten times greater performance than previous generations. Internal test results also showed that a 48-core, fourth-generation Xeon delivered four times better performance than a 48-core AMD EPYC for a range of AI imaging and language benchmarks.

With Granite Rapids and Sierra Forest, Intel will address current limitations for AI and high-performance computing workloads such as memory bandwidth, with 1.5TB memory bandwidth capacity, and 83% peak bandwidth increases over current generations.

Seperately, Intel is also focusing development of GPU and FPGAs (field programmable gate arrays) to meet the demands for large language model training, largely through its Intel Max and Gaudi chips.

It stated that Gaudi 2 has demonstrated two times higher deep learning inference and training performance than the most popular GPUs.

Training on this level is key for large language models (LLM), and demand has risen since the meteoric rise in generative AI models such as ChatGPT.

Around 15 FPGA products will be brought out this calendar year, which will add to Intel’s compute product range, including for deep learning, artificial intelligence, and other high-performance computing needs.

Over time, Intel intends to draw together its GPU and Gaudi AI accelerator portfolios to allow developers to run software to run across architectures.

A plan for an open AI ecosystem

In addition to its achievements and plans for hardware, the firm said it aims to capture and democratise the AI market through software development and collaboration.

With 6.2 million active developers in its community, and 64% of AI developers using Intel tools, its ecosystem already has strong foundations for further AI development.

Intel cited its recent work with Hugging Face, enabling the 176 billion-parameter LLM BLOOMZ through its Gaudi2 architecture. This is a refined version of BLOOM, a text model that can process 46 languages and 13 programming languages, and is also available in a lightweight 7-billion-parameter model.

“For the 176-billion-parameter checkpoint, Gaudi2 is 1.2 times faster than A100 80GB,” wrote Régis Pierrard, machine learning engineer at Hugging Face.

“Smaller checkpoints present interesting results too. Gaudi2 is 3x faster than A100 for BLOOMZ-7B! It is also interesting to note that it manages to benefit from model parallelism whereas A100 is faster on a single device.

Hugging Face noted that the first-generation Gaudi accelerator also offers a better price proposition than A100, with a Gaudi AWS instance costing $13 per hour in comparison to Nvidia’s $30 per hour.

Intel did not provide benchmarks for Gaudi performance next to an H100, the successor to the A100 which is part of the reason big tech is choosing Nvidia for AI.

But lining up Nvidia’s GPUs - long considered best in market - against its own shows Intel is confident that it can deliver and exceed shareholder expectations when it comes to market dominance by 2025.

As part of its Hugging Face collaboration, fourth generation Xeon was used to improve the speed of the open source image generation model Stable Diffusion by more than three times as part of its work with Hugging Face.

The company affirmed its commitment to keep contributing upstream software optimisations to frameworks like TensorFlow and PyTorch, as one of the top three contributors to the latter.

To further open the AI ecosystem, Intel is adding more features to oneAPI, its cross-architecture programming model that offers an alternative to Nvidia’s CUDA software layer.

One of these improves access to SYCL, an open source, royalty-free programming model based in C++ that is heavily used to access hardware accelerators.

Intel’s SYCLomatic can be used to migrate CUDA source code automatically, freeing programmers from time constraints that could otherwise lock them into Nvidia’s architecture.

“We believe that the industry will benefit from an open, standardised programming language that everyone can contribute to, collaborate on, and which is not locked into a particular vendor so it can evolve organically based on its community and public requirements,” said Greg Lavender, CTO and GM of the software and advanced technology group at Intel.

“The desire for an open, multi-vendor, multi-architectural alternative to CUDA is not diminishing. Fundamentally, we believe that innovation will flourish the most in an open field, rather than in the shadows of a walled garden.”

Generative AI could play an important role for organisations looking to prevent cyber attacks such as phishing in the future.

Large language models (LLMs) used by generative AIs such as ChatGPT and Bard could prove effective at learning the language styles of an organisation's staff and be deployed to detect unusual activity coming from their accounts, such as the text in an email.

Kunal Anand, CTO at cyber security firm Imperva, told ITPro that the “cat and mouse game” of cyber security would be enhanced by AI, but that businesses will have to carefully consider what they want used as training data.

Cyber security systems could become more effective if they embedded LLMs that were fed data offering company-by-company context, although such models are largely in the hands of hyperscalers only.

Citing the anti-phishing potential for LLMs in security, Anand said the technology could prove more effective than existing automated security systems due to its capacity for content analysis.

If the technology were ever deployed at scale in the security industry, it's likely that it would augment existing systems, either as standalone products or as features added to unified solutions.

“I think there's going to be interesting use cases where we will use AI, but it will be in conjunction with a signature base system, it will be in conjunction with a logical base system.

"Whether that's a positive security model or a negative security model, I think this is just going to be another layer on top of those things.”

Anand noted that company-specific LLMs could also be a way for large firms to avoid the unwanted collection of valuable data such as source code during the course of training the AI.

Companies such as Google and OpenAI that are producing these AI systems as of now aren't forthcoming with how their tools are collecting or storing the data fed into them, raising questions about safe use in the enterprise.

The potential privacy issues involved with using the tools manifested just last week as ChatGPT was found to leak partial chat histories with other users.

Citing a recent conversation held with employees at an unnamed large company, that had been using GPT to generate client code for internal APIs and microservices, Anand suggested that organisations are already putting too much sensitive data at risk.

“I said, ‘okay, so let me get the straight. You're using some third-party solution, and you have no idea how they’re storing this data, and you're asking it to build you a proprietary application internally using your proprietary schema, that represents your proprietary APIs? You can see the problem right?’

"And they said 'yeah, I don't think we should do that anymore', and I replied 'yeah, I don't think you should do that anymore either'."

“I absolutely believe that from an enterprise perspective, companies from let's say data, security data governance perspectives will probably urge that they bring these generative AI capabilities in-house," Anand said.

"That way they can use their proprietary data and mix it with it.”

Generative AI in malware development

On the other side of the threat landscape, there are already concerns that generative AI can be used to dramatically improve the complexity of malware developed by threat actors.

In January, threat researchers at CyberArk Labs developed polymorphic malware using ChatGPT, a demonstration of the potential threat that generative AI poses to traditional security countermeasures.

Recent calls for leaked AI models to be stored on Bitcoin could exacerbate this potential misuse of LLMs, as a route through which threat actors could anonymously download full training sets that could then be run on home systems. Anand acknowledged that threat actors are already using GPT models to produce malicious code.

“People developing novel attack payloads using these attack tools, and they're asking very open-ended questions of GPT to go in and generate a unique payload that is, you know, something that embeds cross-site scripting or SQL injection - some OWASP top-ten issue and those typically will get sussed out by firewalls in general.”

Generative AI could be used to improve a technique known as ‘fuzzing’, which involves developing an automated script to flood a system with randomised inputs to expose potential vulnerabilities.

Fuzzing tools have been used to expose flaws in popular software such as Word and Acrobat, and generative AI could improve the accuracy with which fuzzing software can iterate on attack results to discover flaws.

“If I try to launch an attack and you block it, I can then use that signal to train my AI, note that that was not a valid payload and try again,” Anand noted.

"And then I can keep mutating over and over and over again, until I find the boundary conditions in your language model.”

Despite the threat of code generation, the use of LLMs for creating malicious text is a bigger issue at the moment.

This is due in part to the complicated nature of programming, with code required to undergo a validity check before being run in a way that prose is not.

Large language models (LLMs) could soon be stored on Bitcoin as a way to distribute publicly-trained LLMs securely and anonymously.

An AI developer has issued a call to action in which they have urged AI developers to disseminate Alpaca and LLaMA, two LLMs derived from leaked Meta training data, via BitTorrent and Bitcoin.

The process aims to democratise generative AI and would involve users publishing AI model data via a torrent and then inscribing it onto the Bitcoin blockchain through a process known as Ordinals, which allows data to be stored on top of Bitcoin.

Once the files have been successfully uploaded, they could become accessible to any user around the world in a decentralised manner.

Any user that had the Bitcoin address of the holder of the model would be able to query Bitcoin for the holder’s transaction history, and use the goatfile to access the torrented models via the bundled magnet URIs.

Torrent files are linked via magnet URIs, which are used to identify files by cryptographic hashing as opposed to by location, and these have been bundled together into a ‘goatfile’ - YAML text that can be inscribed onto $10-20 (£8-16) of Bitcoin - in order to anonymously direct users to the leaked LLaMA files.

The developer tweeted the instructions to widespread interest on 23 March. They noted that once one person has followed all the steps, the models will be forever accessible online.

In this way, blockchain technology could be used to distribute publicly-trained LLMs securely and anonymously.

Meta’s LLaMA LLM was leaked online in early March, and quickly became available via torrent as well as through the imageboard website 4chan.

The model, which Meta made available on a request-by-request basis, is a versatile build capable of processing 20 languages and performed well against AI benchmarks.

A key benefit of LLaMA in comparison to competing models such as LaMDA or GPT-3 and GPT-4 is its small size. The model is available in sizes of 7 billion, 13 billion, 33 billion, and 65 billion parameters, and users can run the smallest of these on consumer hardware as affordable as a single GPU.

A debate over democratised AI

While generative AI has long been the remit of academia, the past year has seen rapid expansion into the field by private companies such as OpenAI, Microsoft, Google, Meta, and Nvidia.

The development of generative AI has fallen largely into the hands of big tech companies, partly due to the need for hyperscaler infrastructure in order to adequately train models containing many billions of parameters.

There are some exceptions to this trend. AWS and Hugging Face have partnered to ‘democratise’ ML and AI, and through LLaMA Meta intended to allow widespread development and implementation of LLMs for approved parties at no cost.

Localising models to AWS, Azure, or Google Cloud architecture will allow developers to complete vast training in short timeframes at manageable cost.

But open source developers, some of whom regard any limitations on software access as an existential threat to the standard, oppose the limitations on access being enforced by such companies.

In the face of this, the LLaMA leak on 4chan radically altered the landscape for the public availability of LLMs. Although the model is far from plug-and-play in its raw form, the fact that it can be run locally means that the model can be used to easily power any number of systems.

There are concerns that the lightweight and powerful model could be used for malicious purposes, and with access achievable through the blockchain it will not be possible to identify which threat actors have obtained the model.

“Anyone can fine-tune this model for anything they want now,” tweeted information security consultant Jeffrey Ladish.

“Fine tune it on 4chan and get endless racist trash. Want a model that constantly tries to gaslight you in subtle ways? Should be achievable. Phishing, scams, and spam are my immediate concerns, but we'll see…”

Just how much of a benefit or a risk some of these open source systems could be has yet to be seen.

Stanford’s Alpaca model, its own academic iteration built on the 7 billion-parameter version of LLaMA, demonstrated that LLaMA is a viable foundation for competing with GPT-3.

It also showed that refining LLaMA against questions generated by GPT-4 could produce proportional improvements as demonstrated in the said model.

Nvidia has announced a series of big tech partnerships this week, including collaborations with AWS and Adobe to provide cloud infrastructure and hardware for training proprietary AI models.

AWS will work with Nvidia to build scalable infrastructure with the express purpose of training large language models (LLMs) for use in generative AI systems. The cloud giant’s P5 elastic compute instances run on Nvidia hardware and provide support for the largest and most complex LLMs.

Adobe and Nvidia have also announced that they will co-develop a series of generative AI models for eventual integration within Adobe’s Creative Cloud suite.

Adobe Firefly is one such model that will focus on image and text effect generation. Being trained on Adobe's own licensed content, or copyright-free material, it's tipped to be an ideal solution for businesses looking to shield themselves from potential plagiarism claims.

Why is Nvidia big tech's go-to partner on AI?

Core to the growing number of big tech agreements Nvidia has secured is its growing dominance in the AI ecosystem.

This was bolstered with the announcement of the Nvidia DGX cloud platform, an AI supercomputing service that lets enterprise customers rent Nvidia’s supercomputing servers and workstations through a web interface.

Oracle Cloud Infrastructure will be the first to host Nvidia DGX Cloud and Nvidia has already announced that other high-profile cloud services providers such as Microsoft Azure and Google Cloud will follow suit, with the tipped for adoption next quarter.

Nvidia’s long-running expertise in designing graphics processing units (GPUs) has also made it particularly sought after for the development of AI models.

The firm has said that GPUs are up to 20 times more energy-efficient than central processing units (CPUs) for AI work, and that its H100 GPU is up to 300 times more efficient for training LLMs than CPUs.

Microsoft’s AI supercomputer, which the firm has revealed it made in time to train both OpenAI’s GPT-3.5 and GPT-4 models, makes heavy use of the H100 for computing power.

It's not clear if this is the same ‘AI supercomputer’ Nvdia and Microsoft announced in November, or whether the firms have partnered again on a longer-term project. In either case, Nvidia has a clear role in Microsoft’s plans for AI going forward.

“Nvidia's status in AI is certainly due in large part to the performance of its GPUs, but this performance is not just a result of hardware,” James Sanders, principal analyst cloud and infrastructure at CCS Insight, told ITPro.

“Nvidia's CUDA software framework provides a common platform that allows developers and users to run the same software across different models and successive generations of GPUs, making generational upgrades seamless, as well as making original development substantially easier."

Sanders added that while Nvidia does have competition in the GPU market from AMD and Intel, it benefits from its competitors' imperfect software support. On rival cards, adaptations need to be made in order to for them to support critical models such as Stable Diffusion.

At the enterprise level, Nvidia faces competition from purpose-built AI accelerators, which are technically faster than Nvidia's GPUs, but require "significant effort to adapt as existing models", Sanders said.

These accelerators also can't be found in easily accessible hardware like consumer-grade laptops, which means developers have to solely rely on cloud platforms if they want to experiment.

“You hear a lot about LLMs at the moment, but there are also a lot of other AI models such as computer vision and NLPs, and these need to be supported at both the hardware and software layer," said Bola Rotibi, chief of enterprise research at CCS Insight, to ITPro. "Nvidia is providing that, as well as choice over processors as well as AWS’ Graviton 3 processors.”

Does Nvidia face competition in this space?

AMD is a major competitor for Nvidia in the GPU market, occupying a 12% share compared to Nvidia’s 17% in Q4 2022, according to Statista.

But while the firm’s Instinct line of GPUs is specifically targeted for deep learning, it has not announced as expansive a foray into generative AI as its competitor.

“The reason why Nvidia is actively trying to use AI technology even for applications that can be done without using AI is that Nvidia has installed a large-scale inference accelerator in the GPU,” David Wang, SVP engineering for Radeon at AMD, told 4Gamer in an interview that was machine-translated from Japanese.

“In order to make effective use of it, it seems that they are working on a theme that needs to mobilise many inference accelerators. That’s their GPU strategy, which is great, but I don’t think we should have the same strategy,” he added.

Without similar investments in AI infrastructure, it is possible that AMD will miss its chance to establish a similar footing in the market as Nvidia.

In addition to its hardware pedigree, Nvidia’s growing cloud ecosystem has already attracted a number of high-profile media partners including Getty Images, Shutterstock, and financial services firm Morningstar.

With a rapidly-expanding number of partners and customers, it may already be too late for competitors to catch up.

Given AMD’s reputation as a more affordable GPU alternative, the company could become a mainstay for those looking to train open source LLMs.

However, some notable open source AI companies such as Hugging Face have already announced plans to use AWS’ AI ecosystem, which has already put Nvidia ahead in this space.

Nvidia could still find itself facing competition from new challengers in the market. Raja Koduri, formerly Intel’s VP of architecture, graphics, and software (IAGS) division, has announced his resignation in order to start a firm that will challenge Nvidia’s GPU dominance in both gaming and generative AI.

Intel itself has also released a range of hardware that it has designated for AI workloads, such as its high-performance Max Series chip family.

But, Nvidia's focus on diversifying its product offering for enterprise and cloud more broadly will likely see it continue to dominate for some time.

"Immediately, the most visible part of this is its acquisition of Mellanox in 2020 for $7 billion, which provided the company advanced networking capabilities – as AI workloads (like other enterprise workloads) spread across multiple networked servers, high-speed, low-latency networking is essential for high performance,” said Sanders.

Chinese tech giant Baidu has announced its own generative AI chatbot, known as Ernie Bot, in a press conference that aimed to put the firm on a competitive footing with OpenAI and Microsoft.

Ernie Bot is a large language model (LLM) that has been in development at Baidu for some time. The company will look to integrate Ernie in its popular Xiaodu smart speaker and voice assistant products.

In a livestreamed press conference, Ernie Bot was shown answering facts, drawing connections between search results, and generating images on demand. Unlike OpenAI’s recent GPT-4 demonstration livestream, Baidu relied on pre-recorded videos for its demonstration.

The bot was also shown to struggle with questions it did not understand. Early into the announcement, Baidu CEO Robin Li showed that the chatbot would insist that users rephrase questions it could not parse.

Elements of the bot that require further development didn’t appear to have inspired investor confidence. Shares in the company dropped 10% in the wake of the announcement, and were down 6.4% at market close.

“Our expectations for Ernie Bot are close to ChatGPT or even GPT-4,” said Li, as translated by the press conference’s English interpreter.

“During our initial internal testing, we experienced the capabilities of Ernie bought and we feel that it is not perfect yet. However, we need to launch it today because we have a huge demand in the market because of various product lines including search, AI cloud, autonomous driving, and Xiaodu.”

Analysis: Can Xiaodu compete with OpenAI?

Baidu is serious about its AI goals - the firm stated that it spend ¥23.3 billion (£2.8 billion) on research and development in 2022, equivalent to almost 20% of its revenue for the year.

Within China, there seems little doubt that Ernie Bot and Baidu’s wider AI ecosystem is set for commercial success. 650 companies have publicly agreed to sign up for the AI system already, and Baidu’s search engine dominance - 56% of the local market - gives it an edge that could stamp out local competition.

Its planned integration with the firm's smart device and AI spinoff Xiaodu also gives it a competitive edge. Although there was nothing like the multimodal capabilities of GPT-4 on display in the demonstration, none of the major firms pursuing generative AI have successfully implemented their models as virtual assistants.

But there are still big challenges to overcome if Baidu is to sell Ernie Bot overseas. Li openly admitted that Ernie Bot is not as effective with the English language as it is in Chinese, putting it behind OpenAI’s multilingual GPT-4 model and even ChatGPT, which runs on the slightly older GPT-3.5.

Li stated that Ernie Bot has been announced ahead of competing models by Google, Meta, and Amazon. Name-checking Google here made sense, as the search firm might be the closest Western company to Baidu in terms of size, purpose, and blended focus on AI and search.

Ernie Bot operates using more than 100 billion parameters, putting it somewhere between Meta’s 65-billion-parameter LLaMA model and OpenAI’s 175-billion-parameter GPT-3.

While it isn’t terribly consequential for its overall performance, as the size of a model is only as important as the quality of its training, it does show a lack of scale that could put the model at a disadvantage against the largest models such as Google’s 540-billion-parameter PaLM.

When it’s released, Ernie Bot will perform a similar function to Google’s Bard, in its smart augmentation of the search experience through natural language processing. But with Microsoft’s backing of OpenAI pushing the firm’s products into greater integration across the business world, it’s the Redmond giant that Baidu will likely have to go up against in the short term.

Microsoft has admitted that its own search engine AI, Bing Chat, was powered by GPT-4 all along. This surprise announcement recontextualises the AI search space, with Google’s ‘code red’ response to counter ChatGPT having shown the company unprepared to compete with the company’s previous generation model, let alone GPT-4.

Baidu could still benefit from China's strict regulations on AI. OpenAI's products are currently banned in the country, and government controls on internet services that restrict the kind of data scraping that has bolstered Western AI firms might shield Baidu from significant local competition.

OpenAI has announced the release of GPT-4, the successor to its popular GPT-3 and 3.5 models, and has promised “human-level performance” in a more creative and stable package than ever before.

The new multimodal model can accept both text and images as input, and was stated to be more creative, reliable, and nuanced than its predecessor. It has been shown to process documents, photos, and charts at a similar level to text input and unpick complex context and tone from user inputs.

OpenAI showed that GPT-4 can reliably identify and caption objects in images, and use these as input context, in a number of examples. These include processing information from a chart, translating and solving a French exam question, and identifying what is wrong or humorous in a given image.

In a livestreamed demonstration of GPT-4, president and co-founder of OpenAI Greg Brockman used the model to translate a photo of a sketch he had made of a website into working HTML code.

GPT-4 also offers major accuracy and stability increases relative to GPT-3 and GPT-3.5 results, having scored in the top 10% of test-takers in a simulated bar exam while GPT-3.5 scored in the bottom 10%. In its blog post on the release, OpenAI stated the model has shown “human-level performance on various professional and academic benchmarks”.

https://twitter.com/DrJimFan/status/1635694095460102145

The new model can process longer documents than ever before, with accepted strings exceeding 25,000 words, and has enabled the long-form analysis and aggregation of entire web pages.

In controlled tests, it was also able to complete extensive multiple choice questions in 26 languages, at a comprehension and accuracy level higher than that of GPT-3.5's English output.

This could dramatically improve document automation and customer-facing interactions, as well as the aggregation and translation of foreign-language dark web posts by security companies.

The road to GPT-4

Achieving these results required OpenAI to redesign its deep learning stack from scratch, while Microsoft’s partnership and $10 billion investment helped the firm establish a supercomputer to facilitate a stable training process for the vast model.

The work has also helped set GPT-4 apart from the models that came before it, both in complexity and reliability.

“On first review, GPT-4 does seem like an important advancement over GPT-3,” Bern Elliot, research vice president and distinguished analyst at Gartner, told IT Pro.

“Much of it is very new and so it will take some time to really understand where and how its improvements over GPT-3 will be realised. The first feature that leapt out to me was multimodal capabilities of GPT-4, where the inputs can be both text and images. This offers a range of new use cases. For instance, in visualising information or in identifying and describing content.

“A second capability that stood out was the more advanced multilingual abilities – its improved ability to handle inputs in languages beyond English. In the press release, they state that in 24 of 26 languages tested, GPT-4 outperforms the English-language performance of GPT-3.5. I spoke with a bank in Thailand this week that was struggling with how to leverage ChatGPT in Thai, this feature capabilities may allow them to do this.”

Elliot also highlighted the improvements to model steerability, which could help firms to customise user experiences, and the reduction of overly-confident lies or ‘hallucinations’ in GPT-4 compared to previous models.

Hallucinations are a growing issue with generative AI, with firms fighting to maintain creativity in models without encouraging them to invent information when real facts are necessary.

“As stated above, it will take time to fully understand the use cases where GPT-4 will excel," Elliot said.

"However, as indicated in the above functions, there are clearly some business uses that this version can address that were not addressable by GPT-3. There is also the overall improved capabilities of this more advanced model. While that may not make a difference for all, for demanding use cases this may be important.”

OpenAI has been clear that the model is not without its limitations. Like its predecessor, the majority of data used to train GPT-4 is from before September 2021 limiting the scope of its knowledge, and cannot learn knowledge through repeated exposure.

The developers enlisted the help of 50 IT experts to adversarially test the model and report its weaknesses in order to reduce unwanted outputs seen in GPT-3.5 such as lies and potentially harmful content.

“We spent six months making GPT-4 safer and more aligned,” OpenAI stated on the GPT-4’s product page.

“GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations.”

Subscribers to OpenAI’s paid tier, ChatGPT Plus, have access to GPT-4 as of now, while those seeking ChatGPT API access to GPT-4 have been asked to join a waitlist. OpenAI stated that select developers are being given access to the API as the firm scales capacity for GPT-4.

OpenAI also revealed that it is making Evals, the software framework used to evaluate large models such as GPT-4, open source. Those that submit “high quality” evaluations of OpenAI models through Evals run a chance of receiving GPT-4 API access, in an incentive that will help OpenAI rapidly collect data on the performance and quality of its products.

GPT-4 will cost $0.03 (£0.025) per thousand prompt tokens up to an 8,000 token limit, and $0.05 (£0.049) per thousand up to a 32,000 token limit.

In response to the announcement, Microsoft has revealed that its Bing chatbot, which has helped drive the search engine’s daily traffic above 100 million users for the first time, has been running on GPT-4 all along.

“If you’ve used the new Bing preview at any time in the last five weeks, you’ve already experienced an early version of this powerful model,” wrote Yusuf Mehdi, corporate vice president and consumer chief marketing officer, at Microsoft.

OpenAI has also revealed a number of partner companies that have already adopted GPT-4 into their stacks.

Financial services firm Stripe has used GPT-4 to summarise websites of potential clients, read and explain complex documentation, and detect fraudsters through syntactic analysis.

Morgan Stanley has found GPT-4 an ideal foundation for its internal chatbot that can draw together the sum knowledge of the firm from across its content library including PDFs. The technology has also been implemented as a virtual assistant in Danish app Be My Eyes, which normally pairs blind or low-vision users with volunteers over video calls.

From having described the contents of a fridge and recommending a suitable recipe to successfully having guided a user through a railway journey step-by-step, the model has demonstrated a clear ability at processing visual input. Be My Eyes is currently OpenAI’s only partner for image inputs, with a wider release on the horizon.

It can be tempting to label every new kind of automated technology tool as something that's entirely never-before-seen. This is especially tempting with innovations in the artificial intelligence (AI) sphere. Although ChatGPT, and other forms of generative AI, may be grabbing the headlines, based on the breadth and depth of use cases, text-generating chatbots have been around for a while and have come to define the customer experience.

Currently, interactions between chatbots and humans mostly happen in the customer service space, and these interactions are mixed. While in some cases, chatbots can helpfully surface the answer to your pesky question that's buried deep within a website's FAQs, a lack of nuance and context defines most interactions, and they serve as a frustrating barrier before the inevitable phone call with a human operator.

ChatGPT, meanwhile, offers far more depth, and possibilities in this space, due to the way the system's been engineered. Yes, it’s still a chatbot and will certainly exhibit some flaws that won’t make it suitable for all use cases, but Open AI’s ChatGPT could revolutionise the interactions between people and machines in all kinds of instances.

When were the first chatbots used?

The history of chatbots can be traced all the way back to Alan Turing in the 1950s. After the advent of his Touring Test, the first chatbot that we can confidently identify, as Greek researchers did in a 2020 literature review, was created in 1966. That bot, named ELIZA, combined pattern matching with pre-written responses and was the first attempt at identifying whether a human audience could tell if a bot speaking to them was, in fact, automated. The notion of AI, labelled as such in the mid-fifties by a group led by researcher John McCarthy, would formalise the field as an area of study.

The first time AI was used within a chatbot environment came in the late 1980s. Although we can’t go back in time and speak to Turing or McCarthy, it’s fair to say that neither probably conceived of tools like Siri and Cortana being ubiquitous in our current moment. OpenAI’s entry into the market in 2015 is remarkably recent, with its technologies leaving significant impressions on enterprises in such a short space of time.

How do ChatGPT and other chatbots work?

To understand the difference between ChatGPT and other types of chatbots it’s important to recognise that, while they’re often grouped together for convenience, the term is not ubiquitous. To further complicate things, tech companies might be unable to adequately define the different chatbots they’ve developed, of varying sophistication and application.

For simplicity's sake, let’s first divide chatbots into two camps. The first, rules-based chatbots, are what they say on the tin: an if-this-then-that system governs their behaviour. They are not, as a rule, adapting or thinking for themselves. One example of a rules-based system would be the pop-up you see if you accept Facebook’s invitation to message a business page and are confronted with prompts asking if you’re looking for information like an establishment’s hours or current array of product listings.

The other broad category is AI-powered chatbots. These interfaces rely on natural language processing (NLP) and machine learning rather than pre-programmed chains of responses. These are smart tools to the relatively analogue chatbots of old. From this point, they’re often characterised by either their use case or where they appear. There are also so-called hybrid bots, which use some elements of both rules-based architecture and AI.

What makes ChatGPT so different to chatbots?

There are so many cautionary tales in this space, including Microsoft’s Tay experiment or Meta’s BlenderBot 3. Where previous attempts at generative AI chatbots quickly devolved into chaos, ChatGPT is programmed in such a way as to reduce the sorts of verbal implosions that led to their downfall. In a few words: it is discerning. According to OpenAI, this has been done through reinforcement learning from human feedback, a process in which the bot’s dataset – in this case, an adapted version of sibling product InstructGPT’s – is reinforced via humans.

“We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides – the user and an AI assistant,” the organisation has previously said in blog post. “We gave the trainers access to model-written suggestions to help them compose their responses. We mixed this new dialogue dataset with the InstructGPT dataset, which we transformed into a dialogue format.”

Microsoft's Tay attempt was almost like leaving a chatbot to learn from an angry pack of wolves with problematic intentions – in this case, social media users. ChatGPT’s initial training was more like a university programme with very defined parameters. The bot, which has its own set of ethical concerns, is currently in a research release phase.

What does ChatGPT mean for the future of human-machine interactions?

When comparing ChatGPT with other chatbots, it’s important to make clear distinctions and not get sucked into what can only be described as a fervent hype machine.

The history of both AI and chatbots is long, winding, and complicated. While generative AI is the new and shiny trend, ChatGPT is just one example of a chatbot that people might interact with now and in future years. This is only one example of an AI-powered textual generation tool, for instance, with OpenAI also working on GPT-4, and other models sure to dazzle the market in future.

It’s too early in the release cycle for ChatGPT to say, for sure, how many businesses might adopt it or integrate its technology into their own tools and services, but a handful of companies have already begun exploring integrating this technology. What’s clear, for now, is the global enterprise landscape is on notice regarding the potential impact of generative AI tools like ChatGPT.

Microsoft has announced that Bing has topped 100 million daily active users (DAUs) for the first time, with its new AI chatbot driving previously unseen interest in the platform.

The Redmond giant stated that a third of the platform’s current monthly users are new, and credited its new technology as improving the overall Bing search experience. Of the users enrolled in its Bing Chat preview, around one-third are also using the chatbot every day, with three chats per session.

The past month has seen 45 million ‘chats’, in which users prompt the model with natural language questions and receive a tailored response.

Each user interaction trains the model to provide more helpful answers in future, which makes the expanded daily users helpful for Microsoft’s wider AI efforts in addition to acting as a marker of consumer confidence.

Like ChatGPT and competing models, Microsoft’s model is also capable of generating ‘new’ written content such as blog posts, poetry, or email copy.

The company stated that 15% of user chats have been creative in nature, an indication that this function has already been popular and meets user demand.

In January, Microsoft announced that it would soon be implementing an OpenAI large language model within Bing called ‘Prometheus’, putting its mammoth $10 billion (£8.4 billion) investment in the AI firm to immediate use.

Since then, the firm has put Bing Chat into public testing via a waitlist, announced plans to integrate the chatbot into Edge browser as well, and debuted the technology on the Bing app.

Microsoft stated that since integrating Bing Chat into the app users on this platform have also increased 600%. The firm acknowledged that its search market share is still small - Statista recorded Bing at 9% vs Google's 85% in January 2023 - but that "it feels good to be at the dance".

“Two factors are driving trial and usage,” wrote Yusuf Mehdi, corporate vice president and consumer chief marketing officer at Microsoft.

“One is Microsoft Edge continues to grow in usage as it has done for the last seven quarters based on the quality of our browser. We expect new capabilities, like having Bing search and create in the Edge sidebar, will bolster further growth.

“The second factor driving trial and usage is that our core web search ranking has taken several significant jumps in relevancy due to the introduction of the Prometheus model so our Bing search quality is at an all-time high.”

One of the advantages of Bing Search over current AI models including the adjacent ChatGPT is that the chatbot provides citations for its claims. This allows users to follow up on the output in more detail if they wish, and fact-check claims that the AI makes.

The chat also offers users the choice of three “conversation styles” for its chatbot: ‘creative’, ‘balanced’, and ‘precise’. The aim is to give users control over the degree to which the AI gets balanced facts with creative responses, to give users the choice of generating inventive responses when desired without taking away from the accuracy of responses when pure facts are desired.

In IT Pro’s own tests, the 'creative' mode was more effective at drafting messages, but also more likely to include details that weren't asked for such as the invented phone number in the first example below.

In contrast 'precise' was true to form in providing a more detailed answer for information on LockBit, the ransomware group notable for its recent negotiation with Royal Mail.

Google, which reportedly ‘upended’ internal teams to compete with ChatGPT in December, has unveiled its own generative AI search proposition with Bard.

The company is betting on its billions of users to retain its title as the king of search, but in the wake of a disappointing launch for Bard that saw an AI error tank Google stock by 7%, has also issued broad warnings against the accuracy of generative AI.

With user demand growing by the month, both Google and Microsoft are set to invest significant figures into this burgeoning market.

Other big industry players have also embraced the potential of generative AI, with Salesforce having announced Einstein GPT for CRM this week along with a partnership with OpenAI.

The firm has stated that it seeks to build an ecosystem of AI partners to expand model access to customers, indicating strong long-term interest in the field.

OpenAI has made its API for ChatGPT generally available, based on a cheaper model that allows developers to easily call on the powerful generative AI for in-app usage.

ChatGPT API accesses a model known as ‘GPT-3.5 turbo’, the same used for the ChatGPT web product. Developers can easily interact with it via a simple endpoint and use it for any of the tasks that ChatGPT is capable of undertaking, from within their website or application.

Leaning on its agreement with Microsoft, the model runs on Azure compute infrastructure which connects to user endpoints. This allows it to run independently of server load on the ChatGPT website, offering businesses a dedicated lane for AI processing.

The API will cost firms $0.002 (£0.0017) per 1,000 tokens, a value worth around 750 words, a sum 10 times smaller than other GPT-3.5 models.

OpenAI stated that this was made possible by a 90% cost reduction in ChatGPT since December, without going into details on how this was achieved.

Enterprise customers seeking reliable access to the model can also purchase dedicated instances, in an agreement in which OpenAI will allocate Azure compute infrastructure solely for the customers’ use.

OpenAI has stated that this arrangement may be the most economical for developers expecting requests in excess of 450 million tokens per day, and can be agreed via direct contact with the company.

OpenAI admitted that it has not met its own targets for delivering a stable service since December, but that it is committed to achieving this over time.

“For the past two months our uptime has not met our own expectations nor that of our users,” read the blog post.

“Our engineering team’s top priority is now stability of production use cases - we know that ensuring AI benefits all of humanity requires being a reliable service provider. Please hold us accountable for improved uptime over the upcoming months!”

Data processed through the API is not used for model training or other improvements to its service unless the organisation chooses to opt-in, and OpenAI has shelved its ‘pre-launch review’ policy which had required developers to flag what they used the model for before it could be integrated within their app.

“Data submitted to the OpenAI API is not used for training, and we have a new 30-day retention policy and are open to less on a case-by-case basis,” tweeted Sam Altman, CEO at OpenAI.

“We've also removed our pre-launch review and made our terms of service and usage policies more developer-friendly.”

The latest model, gpt-3.5-turbo-0301, will be supported until June and a new stable release of the gpt-3.5-turbo is expected in April.

Developers will be given the choice to adopt stable models or specific models according to their needs.

Popular apps that have already made use of the ChatGPT API include Shop, the commerce app by Shopify, which has implemented ChatGPT API to provide more accurate in-app searches and personalised product suggestions for users.

Whisper, OpenAI’s speech recognition system launched in September 2022, was also made available via API at $0.006 (£0.005) per minute of transcribed audio. The open-source model is capable of both transcribing audio and translating speech into English, and can process a number of common audio and video file types.

Through the API, developers can leverage the new ‘large-v2’ model for Whisper which brings speed and quality improvements to its output.

The API offering represents another step towards the full monetisation of ChatGPT, a vital task in the wake of reports that OpenAI’s models are too expensive to run without sizeable income. Altman himself described the firm’s costs as “eye-watering” in December 2022.

Much of this funding may have now been secured through Microsoft’s $10 billion investment in OpenAI, which cemented the dominance of both firms within the AI market. ChatGPT has an ever-expanding presence on Azure and the Redmond giant’s decision to integrate GPT-3.5 into Bing and Edge puts it on competitive footing against Google.

Separate from its influential investors, February saw OpenAI launch its paid tier ChatGPT Plus in the US, offering subscribers faster response times, stable access, and priority updates for $20 (£16) per month.

Some have questioned the cost of ChatGPT Plus in the wake of the API announcement, particularly given that the API allows access to the GPT-3.5 large language model (LLM) at a fraction of that price.

“I hope this pricing impacts ChatGPT+,” wrote a user on the Y Combinator forums.

“$20 is equivalent to what, 10,000,000 tokens? At ~750 words/1k tokens, that’s 7.5 million words per month, or roughly 250,000 words per day, 10,416 words per hour, 173 words per minute, every minute, 24/7. I do not have that big of a utilisation need. It’s kind of weird to vastly overpay.”

OpenAI has announced ChatGPT Plus, a paid tier for its famed chatbot which will offer users faster response times as well as access to the tool even during peak hours.

The new tier is being launched through a US-exclusive pilot scheme, and will cost $20 (£16) per month.

OpenAI stated that access will expand to other countries and regions 'soon' and that free access to the free service will continue.

Users who are willing to pay for the service will also receive priority access to the latest updates and features to the service, though the firm has not specified if this would apply to newer models such as the anticipated GPT-4.

OpenAI confirmed it's considering other paid options, including business plans and lower-cost deals. It did not explain what could differentiate these plans from ChatGPT Plus, though the business plan could seek to provide access to GPT models at the scale and volume required by large enterprises.

Paid tiers will help fund the free generative AI service for other users, which OpenAI runs at great cost.

Experts had questioned the profitability of the company and its continued ability to operate in the wake of CEO Sam Altman’s tweet stating “we will have to monetise it somehow at some point; the compute costs are eye-watering”.

Although it does not publicly report operating costs experts such as Tom Goldstein, associate professor in the department of computer science at the University of Maryland, have estimated that the cost of running ChatGPT alone is high.

“I estimate the cost of running ChatGPT is $100,000 (£81,773) per day, or $3 million (£2.4 million) per month,” Goldstein tweeted.

ChatGPT's popularity boom

OpenAI is also working on a ChatGPT API, for which there is currently a waitlist. This could enable developers to use ChatGPT more seamlessly with products and services such as chatbots on company websites, or to aggregate and translate information using ChatGPT’s pre-trained models.

“We launched ChatGPT as a research preview so we could learn more about the system’s strengths and weaknesses and gather user feedback to help us improve upon its limitations,” OpenAI stated in a blog post announcing the pilot.

“Since then, millions of people have given us feedback, we’ve made several important updates and we’ve seen users find value across a range of professional use cases, including drafting and editing content, brainstorming ideas, programming help, and learning new topics.”

Microsoft’s multi-billion dollar deal with OpenAI, which has aligned the companies closely without resulting in an outright acquisition by the Redmond-based giant, is set to rebalance the market for generative AI.

The firm has already announced plans for a general release of ChatGPT on Azure OpenAI, as well as OpenAI’s image generation model DALL·E.

The release of Microsoft Teams Premium has also seen OpenAI’s GPT 3.5 model integrated within the productivity platform.

It is intended to improve quality of life across Teams through features such as automatically-generated meeting notes, and recommended tasks personalised to users.

Reuters reported that the firm is projecting $200 million (£162 million) in revenue across 2023, and $1 billion (£812 million) by 2024.

OpenAI currently has 375 employees, but has dozens of roles open at the time of writing, across a range of specialisms.

Although it has dominated headlines, OpenAI is not the sole player in the field of generative AI. A number of free alternatives for generating text using pre-trained models are available online. which force the firm to distinguish the unique benefits of its platform as it asks users to pay for use.

Insiders claim that Google has ‘upended’ internal teams to compete with ChatGPT, following concerns that the decades-old player in AI could face reputational damage and sacrifice market control if it lost out to the relatively young OpenAI.

CNBC reported that this includes a product codenamed ‘Apprentice Bard’, which will utilise Google’s Language Model for Dialogue Applications (LaMDA) and operates in a similar manner to ChatGPT. It also alleged that Google is internally trialling a variant of its search page which offers conversational answers to search terms.

This could put Google in more direct competition with Microsoft, which is reportedly seeking to integrate ChatGPT within Bing search to improve user experience and provide a richer user experience.

Both companies will have to tackle latent issues with current AI models that allow them to produce incorrect answers with a high degree of confidence, which could affect the accuracy of search results produced directly by a model.

IT Pro has approached OpenAI for comment.

Having cut through to consumers and businesspeople alike, most could answer "what is ChatGPT" on some level, though under the surface there are many more questions about ChatGPT that may need answering. Since its release on November 30 2022, ChatGPT has become a byword for generative AI and its developers OpenAI have increasingly pitched it as relevant for business use.

With the generative AI market set to expand to $109 billion US (approximately £91.4 billion) by 2030 according to Grand View Research, these tools are becoming more and more important in the enterprise landscape. Usage of the tool remains strong no signs of slowing, even as competitors such as Amazon Q, Google Bard and Anthropic’s Claude 2.1 have entered the market.

ChatGPT updates and summary

ChatGPT’s base model was trained on vast quantities of data scraped from across the web, including entire web pages and books. This was then fine-tuned using reinforcement learning from human feedback (RLHF). That process, from the organization’s perspective, was about ensuring responses were as relevant as possible, removing app bias, and reducing untruth in output otherwise known as AI ‘hallucinations’.

“We needed to collect comparison data, which consisted of two or more model responses ranked by quality,” OpenAI says. “To collect this data, we took conversations that AI trainers had with the chatbot. We randomly selected a model-written message, sampled several alternative completions, and had AI trainers rank them. Using these reward models, we can fine-tune the model using Proximal Policy Optimization.”

GPT models are used to break natural language prompts down into  representations of the words themselves known as vectors. This is used to process the context of the sentence and map the connections between the words it contains, which are used to inform output. In simple terms, the model predicts the most likely words to come next in the sentence based on its training.

Since its launch, OpenAI has released more versions of ChatGPT aimed at business and enterprise use. ChatGPT Enterprise is a plan for the tool that comes with AES-256 encryption for data passed to it, as well as more advanced data analysis capabilities.  

ChatGPT’s latest updates benefitted businesses greatly, bringing live data from the internet through an integration with Bing and OpenAI continues to add features and enterprise uses to the tool.

For one interesting example of an area in which ChatGPT can flourish, we can look at research coming out of Drexel University. A study released in December 2022 found the same natural language processing (NLP) techniques ChatGPT uses can be deployed to identify Alzheimer’s patients.

At CES 2023, Microsoft announced a $10 billion investment into OpenAI, with the hyperscaler having since made OpenAI’s GPT-4 a core part of its Copilot range of AI assistants including Copilot for Microsoft 365 and Copilot Studio. 

For OpenAI’s part, the company has released detailed pricing outlines for its models. Those wishing to run inference on GPT-3.5 Turbo can expect to pay $00.10 per 1k input tokens, while those who want to access the power of GPT-4 will pay the higher price of $0.03 per 1k tokens.

Corporate objections to ChatGPT have also been raised. In May 2023, Apple banned ChatGPT within its offices citing concerns that employees could input sensitive data which could then be leaked and 80% of C-suite executives share these concerns.

Clearly there are still bugs to work out. Research from Purdue University found that more than 50% of ChatGPT’s programming answers were incorrect and warned against overreliance on the chatbot when it came to assessing code.

ChatGPT has an incredibly compelling sales pitch, but it’s still only the beginning for generative AI. Questions around ethical AI will continue to dominate the conversation, especially as the underlying technology evolves and AI-linked job cuts become a risk. A recent study by Boston Consulting Group (BCG) suggested that overreliance on ChatGPT could harm worker performance and these concerns will only become more apparent and relevant as the technology beds in.

The terms machine learning and deep learning can seem interchangeable to most people, but they aren’t. Both considered subdivisions within the world of artificial intelligence (AI), the two have many differences, especially in the architecture and use cases.

Machine learning, for instance, uses structured data and algorithms to train models, with the more data at disposal generally equating with more accurate and better trained models. The idea is to eliminate the need for human intervention. Deep learning, on the other hand, is a subset of machine learning and uses neural networks to imitate the way humans think, meaning the systems designed require even less human intervention.

Differentiating the two, in this way, is crucial to AI research and practical application of both, particularly as businesses attempt to integrate such technologies into their core processes, and recruit for skilled individuals to fill technical roles.

What is machine learning?

The chances are that you’ve probably used an application or system built on machine learning. Whether you’ve interacted with a chatbot, utilized predictive text, or gone to watch a show after Netflix recommended it to you, machine learning was likely at the core of these systems. Machine learning is a subset of AI, and a blanket term used to define machines that learn from datasets. 

Using structured data that comes in the form of text, images, numbers, financial transactions, and many other things, machine learning can replicate the process of human learning. Collected data is used as training material to direct the machine learning model. Theoretically, the greater the volume of data that is used, the higher the quality of the model. Machine learning is all about allowing computers to self-program via training datasets and infrequent human interventions. 

Supervised learning, semi-supervised learning, unsupervised learning, and reinforcement learning are all differing strands of machine learning processes.

The first of these techniques, supervised learning, involves machine learning scientists feeding labeled training data into algorithms to clearly define variables. This is so that the algorithm can start to understand connections. In contrast, unsupervised learning uses unlabelled data and allows the algorithms to actively search for relationships and connections. Acting as the logical midpoint between these processes, semi-supervised learning aids the model’s own comprehension of the data. Reinforcement learning, on the other hand, works by letting a machine complete a set of decisions for the purpose of achieving an objective in an unknown environment. 

What is deep learning?

A subset of machine learning, deep learning deploys systems of artificial neural networks to mimic the cognitive operations of the human mind. 

A lack of sufficient compute power has, until now, created barriers to neural network learning capabilities. Huge strides in big data analytics have changed the landscape significantly, with larger and more complex neural networks now able to take form. This means that machines can now understand, learn, and react to complex scenarios quicker than human beings.

These neural networks are constructed in layers and designed to enable the transmission of data from node to node, much like neurons in the brain. Vast datasets are required to build these models but, once they’ve been constructed, they can give users instant results with little needed in the way of human intervention. There are many, varied ways in which deep learning can be performed.

Convolutional Neural Networks (CNNs): These comprise multiple layers and are mostly used for image processing and object detection.

Recurrent Neural Networks (RNNs): These are types of artificial neural network that use sequential data or time series data. They are frequently used in problems, such as language translation, natural language processing (NLP), speech recognition, and image captioning.

Long Short-Term Memory Networks (LSTMs): These are types of Recurrent Neural Network (RNN) that can learn and remember long-term dependencies. They can be useful for complex problem domains like machine translation, speech recognition, and more.

Generative Adversarial Networks (GANs): These are generative deep learning algorithms that produce new data instances that look like the training data. It comprises two parts; a generator, which learns to generate false data, and a discriminator, which learns from that fake information. These networks have been used to produce fake images of people who have never existed as well as new and unique music.

Radial Basis Function Networks (RBFNs): These networks have an input layer, a hidden layer, and an output layer and are typically used for classification, regression, and time-series predictions.

Multilayer Perceptrons (MLPs): These are a type of feedforward (this means information moves only forward in the network) neural networks. These have an input layer and an output layer that are fully connected. There may also be hidden layers. These are used in speech-recognition, image-recognition, and machine-translation software.

Deep Belief Networks (DBNs): This looks like another feedforward neural network with hidden layers, but isn’t. These are a sequence of restricted boltzmann machines which are sequentially connected. These are used to identify, gather and generate images, video sequences and motion-capture data.

What are the major differences between machine learning and deep learning?

Despite the frequent confusion about their similarities, deep learning is very much a subset of machine learning. Deep learning, however, is differentiated from its counterpart by the data types it interacts with and the ways in which it can learn. 

Machine learning uses structured, labelled data to predict outcomes. This means a machine learning model’s input data defines specific features and is organised into tables. While it gets progressively better at carrying out the task in hand, there still requires there to be a human to intervene at points to ensure the model is working in the required way. In other words, if the predictions are not accurate, an engineer will make any adjustments needed to get back on track.

That being said, deep learning systems involve algorithms that can autonomously decide on the accuracy of their predictions. This works via the presence of neural networks in deep learning models. 

Another difference is that where machine learning can use small amounts of data to make predictions, deep learning needs much, much more data to make more accurate predictions.

While machine learning needs little time to train – typically a few seconds to a few hours – deep learning takes far longer as the algorithms used here involve many layers.

Outputs also differ between the two. Machine learning tends to output numerical values, such as a score or classification, while deep learning can output in multiple formats, such as text, scores, or even sounds and images.

What are the different uses and applications of machine learning vs deep learning?

Machine learning is already in use in a variety of areas that are considered part of day-to-day life, including on social media, on email platforms and, as mentioned, on streaming services like Netflix. These types of applications lend themselves well to machine learning, because they’re relatively simple and don’t require vast amounts of computational power to process complicated decision-making.

Among some of the more complex uses of machine learning are computer vision, such as facial recognition, where technology can be used to recognise people in crowded areas. Handwriting recognition, too, can be used to identify an individual from documents that are scanned en masse. This would apply, for example, to academic examinations, police records, and so on. Speech recognition, meanwhile, such as those used in voice assistants are another application of machine learning.

Because of the nature of deep learning, on the other hand, this technology allows for far more complex decision-making, and near-fully autonomous systems, including robotics and autonomous vehicles.

Deep learning also has its uses in image recognition, where massive amounts of data is ingested and used to help the model tag, index, and annotate images. Such models are currently in use for generating art, in systems like DALL·E. Similarly to machine learning, deep learning can be used in virtual assistants, in chat bots, and even in image colorisation. Deep learning has also had a particularly exciting impact in the field of medicine, such as in the development of personalised medicines created for somebody’s unique genome.

OpenAI has made a new version of its GPT-3 AI language model available that promises to be better at following user intentions while also producing results that are more truthful and less toxic.

The Open AI API is powered by GPT-3 language models that can be used to perform natural language tasks using carefully engineered text prompts. However, the models can also produce outputs that are untruthful, toxic, or reflect harmful sentiments.

The organisation's AI models have been criticised in the past for a range of shortcomings, including racism against specific genders and religions. The organisation once called GPT-3 too dangerous to make public, due to the API being able to create fake news stories by taking cues from the eight million web pages it had scanned to learn about language.

The organisation said this is partly because GPT-3 is trained to predict the next word on a large dataset of Internet text instead of safely performing the language tasks the user wants.

To make its models safer, and more aligned with users, OpenAI used a technique known as reinforcement learning from human feedback (RLHF), using human helpers called labelers to assist the AI in its learning.

“On prompts submitted by our customers to the API, our labelers provide demonstrations of the desired model behavior, and rank several outputs from our models. We then use this data to fine-tune GPT-3,” said the company.

It found the resulting models are much better at following instructions than the GPT-3. They also make up facts less often and show small decreases in toxicity. The organisation’s labelers prefer outputs from its new 1.3B InstructGPT model over outputs from its 175B GPT-3 model, despite having over 100x fewer parameters.

These InstructGPT models have been in beta on the API for over a year and are now the default language models accessible on OpenAI’s API.

“We believe that fine-tuning language models with humans in the loop is a powerful tool for improving their safety and reliability, and we will continue to push in this direction,” the organisation explained.

However, OpenAI outlined that there are some limitations to this model too. The InstructGPT models, for example, are far from fully aligned or fully safe, meaning they still generate toxic outputs, make up facts, or generate sexual and violent content without explicit prompting.

It said that to support the safety of its API, it will continue to review potential applications before they go live, provide content filters for detecting unsafe completions, and monitor for misuse.

OpenAI also highlighted that in many cases, aligning to the average labeler preference may not be desirable. The example it gave is that when generating text that disproportionately affects a minority group, the preferences of that group should be weighted more heavily.

“Right now, InstructGPT is trained to follow instructions in English; thus, it is biased towards the cultural values of English-speaking people,” it said. “We are conducting research into understanding the differences and disagreements between labelers’ preferences so we can condition our models on the values of more specific populations.”

Researchers have discovered a new method of attack against AI systems that aims to clog up a network and slow down processing, in a style similar to that of a denial of service attack.

In a paper being presented at the International Conference on Learning Representation, researchers from the Maryland Cybersecurity Center have outlined how deep neural networks can be tricked by adding more "noise" to their inputs, as reported by MIT Technology Review.

It specifically targets the growing adoption of input-adaptive multi-exit neural networks, which are designed to reduce carbon footprint by passing images through just one neural layer to see if the necessary threshold to accurately report what the image contains has been achieved.

In a traditional neural network, the image would be passed through every layer before a conclusion is drawn, often making it unsuitable for smart devices or similar technology that requires quick answers using low energy consumption.

The researchers found that by simply adding more complication to images, such as slight background noise, poor lighting, or small objects that obscure the main subject, the input-adaptive model views these images as being more difficult to analyse and assigns more computational resources as a result.

The researchers experimented with a scenario whereby hackers had full information about the neural network and found it could be used to max out its energy stores. However, even when the simulation assumed attackers had only limited information about the network, they were still able to slow down processing and increase energy consumption by as much as 80%.

What's more, these attacks transfer well across different types of neural networks, according to the researchers, who also warned that an attack used for one image classification system is enough to disrupt many others.

Professor Tudor Dumitraş, the project's lead researcher, said that more work was needed to understand the extent to which this kind of threat could create damage.

"What's important to me is to bring to people's attention the fact that this is a new threat model, and these kinds of attacks can be done," Dumitraş said.

Microsoft researchers are testing a new feature for Teams that aims to provide speakers on calls with a near-real-time assessment of the moods and reactions of their audience.

The 'AffectiveSpotflight' is said to be built from a type of facial recognition algorithm that uses a neural network to capture and assess the expressions of call participants, monitoring for changes in emotions such as happiness, sadness and surprise.

The software is being developed by researchers across a number of Microsoft facilities in Redmond, Boston and Cambridge, MA, with findings expected to be revealed at Japan's CHI Conference on Human Factors in Computing Systems in May.

The system is said to be able to spot subtle movements, such as the shake of a head, a furrowed brow, and even a raised eyebrow. Each of these is then rated between 0 and 1, with positive emotions scoring higher. The person with the highest score is highlighted to the presenter, every 15 seconds.

The facial expressions of the participants are also matched to datasets in Microsoft's Convolutional Neural Network (CNN), which has expression categories for anger, disgust, fear, happiness, sadness, surprise, and neutral.

"Public speaking is often regarded as one of the most stressful daily activities and is heavily influenced by audience responses to the presenter," the research states. "In fact, studies that seek to reliably induce acute stress on people often involve giving a presentation in front of a neutral-looking audience (a.k.a., Trier social stress test). While research on audience responses in online settings is still nascent, there is prior work considering the impact of in-person audience responses, especially in the context of alleviating public speaking anxiety."

The feature isn't available on Microsoft Teams as yet, but it is very much in keeping with recent updates to the platform that focuses on wellbeing and combating so-called 'video call fatigue'.

However, this may be seen as a somewhat overly technical solution to a problem that is fairly easily solved with feedback, and it isn't difficult to imagine how this feature could create further anxiety as it tries to reduce it.

Researchers at MIT say they have developed a flexible algorithm that can change its underlying equations to continuously adapt to new inputs of data.

The "liquid" algorithm is said to be a new type of neural network that learns during tasks rather than just in its initial training phase.

It's hoped this new approach could revolutionise technology that relies on decision-making protocols where the data changes over time, or in unpredictable environments, such as medical diagnosis or autonomous driving.

The research will be presented at the AAAI Conference, an artificial intelligence event taking place in Vancouver, Canada, in February.

"This is a way forward for the future of robot control, natural language processing, video processing - any form of time series data processing," says Ramin Hasani, the study's lead author. "The potential is really significant."

Most neural networks have fixed behaviour and they typically don't adjust all that well to changes in incoming data streams. For example, the crash of an Uber autonomous vehicle in 2018 that resulted in the death of Elaine Herzberg, considered the first fatality involving the technology, was said to have been caused by the system being unable to identify the shape of a pedestrian when they were walking alongside a bicycle.

The neural network designed by Hasani has the potential to avoid these issues by using a set of differential equations as the base of its algorithm, potentially creating a more fluid type of machine learning. The idea is inspired by the microscopic nematode, Caenorhabditis (C) elegans, which has only 302 neurons in its nervous system. Hasani said they can still "generate unexpectedly complex dynamics".

Similarly, Hasani and his team used equations that allowed the parameters of his neural network to change over time. These are essentially a nested set of differential equations that change the representation of the neuron, creating a small number of highly "expressive" ones, according to Hasani.

"We have a provably more expressive neural network that is inspired by nature," Hasani said. "But this is just the beginning of the process. The obvious question is how do you extend this? We think this kind of network could be a key element of future intelligence systems."

Police Scotland has launched a new neural network-driven drone programme to help find people who have gone missing.

The technology onboard the fleet of drones, such as advanced cameras and AI-powered software, will allow pilots to determine whether something that appears on a screen is, in fact, a person.

When a camera is situated at great heights, it becomes difficult for the human eye to discern a 'speck' on the image from a silhouette of a missing person.

The advanced imagery technology - which is the result of a collaboration between Police Scotland, Thales and the University of the West of Scotland - can help identify missing people from up to 150 metres away. The software it uses to differentiate meaningless 'specks' from animals, vehicles or real people is light enough to be run on a smartphone and uses machine learning to continually increase its accuracy.

"The drone itself has very special sensors on it," said Insp Nicholas Whyte, of Police Scotland's air support unit to the BBC.

"There's a very highly-powered optical camera which can allow us to see things quite clearly from a good height. Also, there's a thermal imaging sensor which detects heat.

"We're there to find people. People who need our help or people who are lost."

Growing adoption

The technology being deployed by Police Scotland is the latest in a line of deployments from emergency services across Britain in recent months.

The UK's Maritime and Coastguard Agency (MCA) announced in August that it was seeking applicants for its 990,000 drone contract to help search for missing persons up to 10km away from shore.

The technology would allow emergency responders to safely and remotely explore regions that would be difficult to investigate due to poor conditions such as low light, wind or fog.

The UK government also awarded funding to a variety of startups looking to use drone technology for more than cool cinematic shots and airport disruptions.

Among the successful applicants was Loughborough University which has developed technology to detect chemical hazards. Others were developing technology to detect gas leaks and assist decontamination missions.

Ordnance Survey also said earlier this year that it would start using solar-powered drone technology to collect better images of Earth than was previously possible.

The 145kg drone named Astigan will be able to fly for 90 days at 67,000 feet without stopping and will capture images which conventional aerial image technology would struggle to do.

It is admittedly quite late in the game to consider the advantages of AI given how commonly used it already is. Nevertheless, it's an important technology to ponder and debate as it will inevitably affect all life on the planet.  

From minuscule algorithms that offer a little convenience for everyday life to international defense systems, AI is already impacting us, often in ways we cannot see. What you see on your social media feeds, how we predict the weather, and how we manage our staff members and workflows, are all now common use cases of AI.   

While we have a thriving AI industry and seemingly a new use case popping up daily, we still have only scratched the surface of what this technology has to offer. Before adopting it, however, it's always worth considering the advantages and disadvantages of AI.

What are the advantages AI?

The advantages of AI are currently being talked about, to death, by every tech company that has a product to ship. You'll have noticed 'AI' being tagged on just about everything. It's in our laptops, our collaboration software, and even our kettles. But what are the actual benefits of artificial intelligence? 

Improved efficiency

This is arguably where the business interest will be, particularly when it comes to their IT stacks and their workforce. Things working more efficiently will lead to improved performance and cost savings. Automation and generative AI tools are coming out in mass and getting both more affordable and more advanced as we speak. 

A key element here is data; Copilots, such as Microsoft's, run on web and user data, analyzing how, when, and why we use our devices and software and offer us shortcuts and simple directions. For a more business example, SAP's Joule Copilot takes in SAP customer data – essentially pure b2b information – to improve how its customers use ERP and financial platforms. That process basically explains most AI efficiencies, data feed into an algorithm to find better, faster, and simpler ways of working. 

Reducing human error

Another efficiency is the reduction of errors – human ones. We are all prone to the odd laps of concentration and we also simply get things wrong. AI-based machines carrying out specific tasks don't – if they do, you'll likely find that's also some kind of human error.  

This is why Amazon and Ocado use robots in fulfillment centers, and why financial systems, such as those that evaluate credit, are based on algorithms. The reasoning here is volume; you can do more if it's automated, but you're also not increasing the risk of error. 

Advancing AI

In some cases, by using AI you are, by default, also improving AI. Maybe only by a little bit, but it is still learning and developing with the data you create. When your business uses a Copilot, said Copilot will have been designed to learn from its users. It will essentially tailor its service towards the users. So, for example, SAP's Joule will help you navigate Concur when filling out business expense applications and gradually figure out your preferences and those of your company.

If we take that example and apply it to a more grander use case, such as quantum computing, or medical diagnostics, AI will only get faster and more efficient as we apply it. This will be dependent on privacy policies, regulations, and the types of systems being used; AI will largely improve through training and fine-tuning. But it is data (our actions) that will help it to advance.

What are the disadvantages of AI?

It’s natural to be fearful of powerful technology. Recent history with data scandals, malware, and social media has made that clear. AI is no different and many of the concerns held by onlookers are, in some areas, justified. But that doesn’t mean great work isn’t being done to mitigate the drawbacks.

In 2016, an industry-wide organisation including five Silicon Valley giants was formed, known as the Partnership on Artificial Intelligence to Benefit People and Society. This body works to promote the fair and ethical development of artificial intelligence technologies that have the potential to bring as much disruption as it will benefit.

Decision-making AI in the workplace

The speed and efficiency of certain AI applications make them appealing to executives looking to find more value across their organizations. 

IBM's Watson has been used to decide if employees are worthy of a pay rise, a bonus, or a promotion by looking at the experience and past projects of employees to indicate the future qualities and skills individuals could bring to the company. 

Decision-making software used in this way has caused some concern. The Trades Union Congress, the federation that represents the majority of trade unions in the UK, called for legislative changes last year to safeguard employees against this kind of technology. It is also recommended that employers consult trade unions before deploying such systems.

"Our prediction is that left unchecked, the use of AI to manage people will also lead to work becoming an increasingly lonely and isolating experience, where the joy of human connection is lost," TUC general secretary Frances O'Grady said.

Job losses

The potential for human job losses is widely regarded as the number one downside to AI, the implementation of which could set in motion a wave of lay-offs as employees struggle to outperform machines.

However, while this scary scenario is often presented as just over the horizon, AI is expected to create more jobs than it takes. The World Economic Forum’s (WEF) 2020 Future of Jobs Report predicts that by 2025, automation will have affected 85 million jobs around the world but 97 million jobs will be created in industries such as artificial intelligence.

“No matter what prediction you believe about jobs and skills, what is bound to be true is heightened intensity and higher frequency of career transitions, especially for those already most vulnerable and marginalized,” stated FutureFit AI CEO Hamoon Ekhtiari.

The report also stated that, although the pandemic has accelerated the automation of many repetitive and dangerous tasks, “around 40% of workers will require reskilling of six months or less”. One area of skills worth developing in time for the AI-based future is data, but soft skills shouldn’t be ignored either. John Whittingdale OBE, former minister of state for media and data, described soft skills as “hugely important”, adding that, “without them, there is the potential for data to be misread or miscommunicated, which can have significant implications for businesses and the decisions they make”.

Human error

Although AI can virtually remove human error from processes, its code is still subject to bias and prejudice. Being largely algorithm-based, the technology can knowingly or unknowingly be coded to discriminate against minorities or fail to cater to groups that its programmers failed to consider.

If security measures are not followed carefully, hackers can exploit AI seeking to collect public data. For example, Microsoft's ill-fated chatbot Tay Tweets had to be taken down after only 16 hours as it had started to tweet racist and inflammatory content driven by input from other Twitter users.

Importantly, Tay Tweets was purposefully fed hateful content in an effort by Twitter and 4chan users to break it. But other examples of AI going astray have come despite the best efforts by its developers.

For example, in 2018 Amazon decided to retire a recruitment algorithm after it was discovered to discriminate against non-male candidates. The AI system was intended to provide hiring recommendations and had been fed ten years of application data to help train its decision-making. However, as the majority of submissions have been handed in by men, the conclusion the AI came to was that men were preferred candidates.

Responsible use of AI

There is a great deal to be positive about when it comes to AI. Any emerging technology that has the power to disrupt the existing structures of individuals and organisations must be assessed for its potential risks.

But being mindful of the downsides does not mean becoming blinkered to the benefits. Indeed, decision-makers have been warned against doing just this, or else risk losing out on the clear improvements that careful use of AI can bring.

"Look at how you are using technology today during critical interactions with customers – business moments – and consider how the value of those moments could be increased. Then apply AI to those points for additional business value," said Whit Andrews, distinguished vice president analyst at Gartner.

"AI projects face unique obstacles due to their scope and popularity, misperceptions about their value, the nature of the data they touch, and cultural concerns. To surmount these hurdles, CIOs should set realistic expectations, identify suitable use cases and create new organisational structures."

Gartner advises that business and IT leaders should endeavour to cut the AI hype away from reality by carefully considering and weighing up the opportunities vs risks. Obsessively focusing on automation, rather than the bigger picture, will only obscure the wider benefits, the analyst firm warns.

In July 2022, the UK government and Alan Turing Institute jointly announced the establishment of the Defence Centre for AI Research (DCAR). Its goal is to develop areas of AI research that are currently proving challenging to implement, such as training without the need for large data sets, AI ethics, and war gaming.

"Everything we love about civilisation is a product of intelligence," said Max Tegmark, president of the Future of Life Institute.

"Amplifying our human intelligence with artificial intelligence has the potential of helping civilisation flourish like never before as long as we manage to keep the technology beneficial."

Organisations can also check if their use of AI systems breaches data protection laws using a risk assessment toolkit launched by The Information Commissioner's Office (ICO). The AI and Data Protection Risk Assessment Toolkit, available in beta, draws upon the regulator's previously published guidance on AI, as well as other publications provided by the Alan Turing Institute.

It contains risk statements that organisations can use while processing personal data to understand the implications this can have for the rights of individuals. Based on an auditing framework developed by the ICO’s internal assurance and investigation teams, the toolkit also provides suggestions for best practices that companies can put in place to manage and mitigate risks.

Machine learning has many broad methods, of which deep learning is perhaps the most well-known. A type of artificial intelligence (AI) technology, the ‘deep’ in deep learning refers to the layered and hierarchical nature of the algorithms it uses to train AI.

The real-world applications of deep learning are broad, but can most commonly be found in use cases such as voice recognition in smart home assistants, and natural language processing models used by AI-driven chatbots.

The ultimate ambition of deep learning is to teach an AI to make its own decisions based on the data it is fed. In the example of a smart assistant, to hard-code all the different types of questions a user could ask a home assistant would be nigh-on impossible and incredibly laborious. The more efficient, and perhaps more time-consuming in the short-term, method of building such an AI would be to train it using deep learning to recognise what it's being asked and how to appropriately respond.

Deep learning models are trained using data – tons of data. The more data that can be fed into a system, the more accurate the decisions it can make. Many deep learning models take inspiration from human biology, using neural networks that arrange analytical nodes in interconnecting pathways, allowing them to replicate the multi-layered (deep) connections found in the human brain.

Is deep learning supervised or unsupervised?

The special draw of deep learning that’s attracting so many to study the field is the potential for both supervised learning and unsupervised learning, although it's the latter that has opened up so many avenues for research.

Supervised deep learning definition

Referring to a training technique using labelled data, supervised learning is one approach researchers often take when training deep learning models. Ahead of training, data is prepped by researchers and labels are applied to both the input data and the output data. This is so the model can directly analyse the relationship between both sets. After the model has a definition of the relationship and once it understands why the relationship exists, it can start to more accurately analyse other bodies of data. 

Such training is referred to as ‘supervised’ owing to the required presence of a human, whose main role is labelling the data so that the model can tell what it's processing. The purposes of this training are often seen in object detection applications and it's considered to be the most popular choice for model training in machine learning.

Deep learning training can also be done in a semi-supervised manner which requires the labelling of just a small proportion of the overall training data set. A much greater amount of unlabelled data is then used simultaneously to analyze a problem. 

Unsupervised deep learning definition

Unsupervised learning refers to the process of using entirely unlabelled data, and feeding it into the deep learning model with a view to training said model to recognise patterns in completely new data. The lack of human intervention creates potential for unforseen determinations to be made by the model, something that forms the basis of a huge body of research today.

There are three main tasks when it comes to unsupervised learning. These are:

• Clustering: Involves looking at unlabelled data and grouping similar data points into groups. This task is often used for training models to compress digital images, for example
• Association: Somewhat similar to clustering, the association approach is used to understand why there is a relationship between data points and how the data points connect rather than just knowing that there is a relationship there
• Dimensionality reduction: Often performed during the pre-processing of data, where researchers attempt to reduce the noise in the training data. This refers to ‘simplifying’ a given data set that has too many features, or dimensions, to accurately analyse. It involves reducing the number of data inputs so that a relationship between input and output data can be established

Examples of deep learning today

Neural networks commonly form the foundation of many advanced machine learning systems in the world today. The self-driving car, for example, has been made possible via deep learning, and the concepts of deep learning are also finding many applications in the defence and aerospace industries. 

Deep learning has its limitations, however, despite its huge potential. Tasks that are more human-like, for example, hold back deep learning capabilities. Deep learning thrives more in the areas of pattern recognition, excelling in understanding the complicated but constant rules of the programming language Go. Even in pattern recognition, a huge amount of data is still required to train deep learning models. 

Tasks of pattern recognition come to the fore most obviously in conversational AI systems, where deep learning constitutes a support network. Multimodal inputs are processed alongside multimodal outputs - think voice recognition capabilities and synthesized voices. Huge companies like Starbucks and Apple are already utilizing these sorts of capabilities, with customers now increasingly presented with opportunities to place orders with voice commands. Other applications include giving consumers the ability to log in to their phones just by looking at the screen. 

At the current stage of development, it does not appear possible for deep learning to perform the same elaborate, adaptive thought processes as humans, however, the technology continues to evolve at quite a rate.

The future of deep learning

Deep learning may not result in killer robots anytime soon, but that isn't to say it won't fundamentally alter aspects of society in other ways.

Google Brain represented how its deep learning AI system was thinking for itself through an experiment involving cats. Though the research group did not specify any experimental parameters for the identification of cats, Google Brain was able to identify images of millions of cats. 

While this may on the surface seem a banal use case, it's not hard to appreciate how this development could be put to a more practical use.

In medicine, deep learning has been found to be on par with human expertise when it comes to interpreting medical images. The study, carried out by the University of Birmingham, may pave the way for AI to play a greater role in the medical field going forwards, easing the strain on resources and allowing doctors to spend more time with patients.

Perhaps the most exciting area where deep learning is being touted as a possible springboard to discovery is in the cosmos. Researchers from ETH Zurich university released a paper in which they employed neural networks to study dark matter. When compared to the Hubble telescope, deep learning was found to deliver 30% more accurate values when breaking down the composites of the universe, apportioning baryonic matter, dark matter and dark energy. The researchers concluded by claiming that deep learning is a promising prospect for cosmological data analysis in the future.

What is certain, is that with funding being poured into AI - the Pentagon allocated nearly $1 billion to AI in 2020 - and specifically deep learning research studies, the technologies' influence will only grow.

At the heart of many artificial intelligence computer systems are artificial neural networks (ANN). These networks have been inspired by the biological arrangements found in a human brain.

Using a structure of connected "neurons", these networks can distinguish numerical arrangements by 'learning' to process certain stimuli and creating assessments without the involvement of humans.

One such practical instance of this is the use of an ANN to identify objects in images. In a system constructed to identify the image of a cat, an ANN will be taught on a data set that comprises images that are labelled “cat”, which can be used as a reference point for any forthcoming analysis. Just as people may learn to recognise a dog based on distinctive features, such as a tail or fur, so too does an ANN, by breaking each image down into their various component parts, such as colour and shapes.

In practical terms, a neural network offers a sorting and classification level that sits on top of your managed data, aiding the clustering and grouping of data based on resemblances. It's possible to produce complex spam filters, algorithms to find fraudulent behaviour and customer relationship tools that precisely measure mood, all using an artificial neural network.

How an artificial neural network works

ANNs draw inspiration from the neurological organization of the human brain. They are constructed using neuron-like computational nodes that converse with each other along channels like the way synapses work. This means the output of one computational node will affect the processing of another.

Neural networks signified an enormous leap in the development of artificial intelligence, which had until then relied on the use of pre-defined processes and regular human intervention to create the desired outcome. An ANN allows the analytical load to be spread across a net of several interconnected layers, each containing interconnected nodes. As information is processed and contextualised, it's then passed along to the next node, and down through the layers. The idea is to allow additional contextual information to be drip-fed into the network to inform processing at every stage.

A basic structure of a single 'hidden' layer neural network

Much like the structure of a fishing net, a single layer of a neural network connects processing nodes together using strands. The vast number of connections enable enhanced communication between these nodes, increasing accuracy and data processing throughput.

ANNs will then pile a number of these layers on top of each other to analyse data, creating an input and output flow of data from the first layer to the last. Although the number of layers will vary depending on the nature of the ANN and its task, the idea is to pass data from one layer to another, with additional contextual information being added as it goes. This deviates slightly from the human brain, which is connected using a 3D matrix, rather than a series of layers.

Like an organic brain, nodes 'fire' across an ANN when they receive specific stimuli, passing the signal over to another node. However, in the case of ANNs, the input signal is defined as a real number, with the output being the sum of the various inputs.

The value of these inputs is dependent on their weighting, which serves to increase or decrease the importance placed on the inputs respective to the task being performed. The goal is to take an arbitrary number of binary inputs and translate them into a single binary output.

A more complex neural network, increasing the sophistication of its processing

Earlier models of neural networks used shallow structures, where only one input and output layer were used. Modern systems now are comprised of an input layer, where data first enters the network, multiple 'hidden' layers, which add complexity to the analysis, and an output layer.

This is where the term 'deep learning' has derived - the 'deep' part specifically referring to any neural network that uses more than one 'hidden' layer.

The evening party example

To explain how a neural network works in practice, let's boil it down to a real-world example.

Imagine you have been invited to a party and you are trying to decide whether to go. It's highly likely that you are weighing up pros and cons and pulling a variety of factors into the decision process. For the sake of this example, we'll pick just three – 'will my friends be going?', 'is the party easy to get to?', 'is the weather going to be good?'

It's possible to model this process using an artificial neural network by translating these considerations into binary numbers. For example, we could assign a binary value to 'weather', namely '1' for clear weather and '0' for severe weather. The same format will be repeated for each deciding factor.

Neural threshold

However, it's not enough to just assign values, as this doesn't help us arrive at a decision. For that, we would need to define a threshold – that is, the point at which the number of positive factors outweighs the number of negative factors.

Based on the binary values, a suitable threshold could be '2' – in other words, you would need two factors to return a '1' before deciding to attend the party. If your friends were to attend the party ('1') and the weather was good ('1'), this would be enough to pass the threshold.

If the weather was bad ('0'), and the party was difficult to get to ('0'), this would not meet the threshold and you would decide against attending the party, even if your friends were to attend ('1').

Neural weighting

Admittedly, that is a very basic example of the fundamentals of a neural network, but hopefully, it helps to highlight the idea of values and thresholds. However, the decision-making process is far more complex than this example, and it's often the case that one factor will be more influential in the decision-making process over another.

To create this variation, we can use 'neural weighting' – this dictates how important a factor's binary value is in relation to other factors by multiplying it by its weighting.

While each of the considerations in our example could sway you one way or another, it's likely that you will place greater importance on one or two of the factors. If you are entirely put off by the idea of leaving the house during a heavy downpour, then the inclement weather will outweigh the other two considerations. In this example, we could place greater importance on the weather value by giving it a higher weighting:

• Weather = w5
• Friends = w2
• Distance = w2

If we imagine that the threshold has now been set to 6, poor weather (a 0 value) would prevent the rest of the inputs from reaching the desired threshold, and therefore the node would not 'fire' (you would decide against going to the party).

Again, it's a simple example, but it provides an overview of how decisions are made based on the weight of evidence provided. If you're to extrapolate this to, say, an image recognition system, the various considerations of whether to attend a party (inputs) would be the dismantled characteristics of a given image, whether that's colour, size, or shapes. A system training to identifying a dog, for example, may give greater weighting to shapes, or colour.

When a neural network is in training, the weights and thresholds are set to random values. These are then continually adjusted as training data is passed through the network until a consistent output is achieved.

Benefits of neural networks

Neural networks are not completely limited by the data they feed on – you can go as far as saying they learn organically. And, as they can generalize inputs, they are highly valuable for pattern recognition systems. 

This includes the ability to find shortcuts when tasked with answering computationally intensive answers and inferring relations between data points. The ANN doesn't have to expect a data source to be unambiguously connected.  

They can also be fault tolerant, in that they have the ability to respond to failures and continue operating despite the error. In effect, it can self-diagnose and debug a network.

If they are scaled over multiple systems, they can route around missing or non-communicative nodes. This is in addition to finding ways around non-functioning parts of the network – it can do this by recreating data by inference to determine where the non-functioning nodes are.

Perhaps the biggest advantage of deep neural networks is the capability to process and cluster unstructured data. This includes images, text, audio, and also numerical data.

Examples of neural networks

Artificial neural networks is an area that is undergoing immense advancement, particularly with newer developments in generative AI. From modeling complex systems to processing ever larger datasets, the evolution of neural networks has been rapid, especially in fields like image processing, speech recognition, the financial and medical sectors, and content creation. 

Image processing is arguably one of the easiest to explain as anyone with a modern smartphone will know. ANNs are used here for image classification and object recognition. 

In the financial sector, ANNs can be used for credit scoring and stock market predictions. Investors and risk managers can use ANNs to process mass volumes of financial data and find telling insights that can inform investment decisions. When it comes to credit scores, it's about data-driven assessments to improve the accuracy of default predictions. Both examples require high-quality data and there is a black-box nature to financial system in that they tend to give results without giving away too much information.

Medical use cases for ANNs mostly deal with detection technologies, such as image recognition for MRI scans. With vast medical datasets, ANNs can enhance the accuracy of diagnostics and potentially catch diseases in their infancy. 

However, ANNs are also helping advancements in drug discovery. Here they are used to speed up and identify potential drug candidates and formulate treatment plans and even financial outlays.  

Another area of growth for ANNs is content creation. However, there are a few issues here, particularly around the use of deep learning models to learn and then 'create' the work of an artist. Essentially this is the power to create the ideas, images, and sounds of someone without needing the actual person. Examples like DALL-E, which is a deep neural network trained on millions of images and volumes of text, can be used to create various artworks - though often it's a mishmash of real-life works. 

Artificial intelligence is creating faces,text, recipes and even CVs. The latest AI and neural network research highlights how far such work has come, but also highlights the complexity of using such models to create something new. We explore five examples of ways AI are creating and the work that goes into powering them.

Imaginary stock photos

Researchers at Nvidia have made people well, photos of them, at least. Building on existing work that uses generative adversarial networks to make images of faces, Nvidia added a system called Transfer, which steals aspects of real faces to enhance fake ones. So it's nabbing a nose from here, eyes from there, and without any human interaction combining them into realistic photos of people who have never existed. The system takes two source photos. Source A will inspire a specific feature, such as gender or age or pose, while source B fills in the rest. You can check out theconvincing examples at thispersondoesnotexist.com or simply look at the lead image on this article.

Such work raises concerns about fake images being too realistic can we trust our eyes anymore? but making the photos not only required neural network skills, but alsoaweek on eight Tesla GPUs to create one set of faces. And spoof versions, such as the uncanny cat-like monsters at thiscatdoesnotexist.com,show how difficult it still is to make realistic fake photos.

Neural network CVs

Why not let a neural network write your CV for you? To be clear, the rsum generator made by Enhancv a CV-constructing site won't spit out a list of your own accomplishments, but instead list work and skills culled from the internet. Why? The company says the aim was to show a wider range of examples of what CVs could be designed to look like. The photos that run alongside many CVs are from Nvidia's project. Head to thisresumedoesnotexist.com to give it a go.

Photo editing the easy way

Photoshopping takes skills. A team of researchers from South Korea has created an algorithm that allows us humans to tell a neural network how we would like to change a photo: switching hairstyles, adding a smile, strengthening their jaw line, even removing sunglasses. Doodle on the image and the neural network understands that you would like the hair to be fluffier or theglasses removed, grabbing its ideas from other images. Not only could it be great for trying out new hairstyles but also for creating spoof images that show a public figure smiling in aserious moment. Welcome to the future.

Flavourful AI

Artificial intelligence isn't just making photos and CVs: it's also making flavours. IBM and spice maker McCormick & Company have teamed up to use AI to generate new food ideas, hoping to cut down the number of iterations a formula goes through before it's ready for home cooks. The AI pulls in market research data, decades of existing formulas, ingredient lists, and key restrictions such as vegetarian or kosher, then invents formulas for food flavourings for sauces, condiments and spice mixes. The results will hit shelves this year, including a flavour mix of spices for Tuscan chicken.

OpenAI's GPT-2

Give this system a prompt and it will spit out coherent sentences on the same theme. Trained on eight million web pages, scraped from sites linked to from Reddit, GPT-2 simply predicts the next word based on previous words an advanced version of autosuggest on phone keyboards. Notonly does it produce seemingly sensible text, but it writes in the correct tone: for example, a story about celebrities takes on a tabloid vibe, while a prompt in the style of a teacher giving homework to a student sparks a school-level essay. That said, it does take a few tries to get a good sample, with the model generating useful text about half the time, and the sentences aren't necessarily factually accurate.

Indeed, OpenAI was so worried that the system could be abused for nefarious uses that it has refused to release the entire model and dataset. While the text generator could be used to build better translation or speech recognition systems, OpenAI is worried it could also be used to make fake news articles and automate spam or improve social media bots. (This article was written by a human, though. Or was it?)

If there's one resource the world isn't going to run out of anytime soon it's data. International analyst firm IDC estimates the Global Datasphere' or the total amount of data stored on computers across the world will grow from 33 zettabytes in 2018 to 175 zettabytes in 2025. Or to put that in a more relatable form, 175 billion of those terabyte hard disks you might find inside one of today's PCs.

That data pool is an enormous resource, but one that's far too big for humans to exploit. Instead, we're going to need to rely on deep learning to make sense of all that data and discover links we don't know even exist yet. The applications of deep learning are, according to Intel's AI Technical Solution Specialist, Walter Riviera, "limitless".

"The coolest application for deep learning is yet to be invented," he says.

So, what is deep learning and why is it so powerful?

Teaching the brain

Deep learning is a subset of machine learning and artificial intelligence. It is is specifically concerned with neural networks an interconnected graph of "digital neurons" inspired by the human brain.

An artificial neural network is comprised of layers of "digital neurons," says Riviera. "The more layers you have, the deeper you go, the more powerful is the algorithm."

There are two key steps in deep learning: training and inference. The first is teaching that virtual brain to do something, the second is deploying that brain to do what it's supposed to do. Riviera says the process is akin to playing a guitar. When you pick up a guitar, you normally have to tune the strings. So you play a chord and see if it matches the sound of the chord you know to be correct. "Unconsciously, you match the emitted sound with the expected one," he says. "Somehow you're measuring the error the difference between the two."

If the two chords don't match, you twiddle the tuning pegs and strum the chord again, repeating the process until the sound from the guitar matches the one you are expecting to get. "It's an iterative process and after a while, once all the chords are playing as expected, you can basically stop iterating and drop the guitar, because is ready to be used," says Riviera. "What song can you play? Whatever, even chords that weren't used to check the tuning steps."

In other words, once you've used a specific dataset to train a neural network, it will be able to evaluate new-unseen data samples independently.

"Once the neural network is giving the expected answers or the error is very close to zero, we have completed the training process. Then it's time to take that artificial brain and deploy it wherever might be relevant for our application: in a big server within a data centre or embedded in portable devices at the edge."

Deep learning as a concept isn't new indeed, the idea has been around for 40 years. What makes it so exciting now is that we finally have all the pieces in place to unlock its potential.

"We had the theory and the research papers, we had all the concepts, but we were missing two important components, which were the data to learn from and the compute power," says Riviera. "Today, we have all these three components theory, data and infrastructures we just need to be creative in order to benefit from the opportunities that deep learning is opening."

Deeper learning

That's not to say that deep learning isn't already being put to amazingly good use.

Any regular commuter will know the sheer fist-thumping frustration of delays and cancelled trains. However, Intel technology is being used to power Resonate's Luminate platform, which helps one British train company better manage more than 2,000 journeys per day.

Small, Intel-powered gateways are placed on the trackside, monitoring the movements of trains across the network. That is married with other critical data, such as timetables, temporary speed restrictions and logs of any faults across the network. By combining all this data and learning from past behaviour, Luminate can forecast where problems might occur on the network and allow managers to simulate revised schedules without disrupting live rail passengers. The system can also make automatic adjustments to short-term schedules, moving trains to where they are most needed.

The results have been startling. On-time arrivals have increased by 9% since the adoption of the system, with 92% of trains now running to schedule.

Perhaps just as annoying as delayed trains is arriving at the supermarket to find the product you went there for is out of stock. Once again, Intel's deep learning technology is being used to avert this costly situation for supermarkets.

The Intel-powered Vispera ShelfSight system has cameras mounted in stores, keeping an eye on the supermarket shelves. Deep-learning algorithms are used to train the system to identify individual products and to spot empty spaces on the shelves, or even products accidentally placed in the wrong areas by staff.

Staff are alerted to shortages using mobile devices, so that shelves can be quickly restocked and lost sales are kept to a minimum. And because all that data is fed back to the cloud, sales models can be adjusted and the chances of future shortages of in-demand products are reduced.

Only the start

Yet, as Riviera said earlier, these applications of deep learning are really only the start. He relays the story of the Italian start-up that is using deep learning to create a system where drones carry human organs from hospital to hospital, eliminating the huge disadvantages of helicopters (too costly) and ambulances (too slow) when it comes to life-critical transplants.

It's not the only life-saving application he can see for the technology, either. "I'd like to see deep learning used to build an autonomous system robots that can go and collect plastic from the oceans," he says. "We would already have that capability, it would just be matter of developing and enabling it."

"The best [use for deep learning] is yet to be invented," he concludes.

Discover more about data innovations at Intel.co.uk

Three artificial intelligence (AI) pioneers have been jointly awarded the Alan Turing Award for 2019, a prize coined 'the Nobel Prize' for AI.

With a hefty 1m prize, the three winners split the pot for their transformative work in neural networking, a technology that heavily underpins the success of AI which works by teaching technology to think in the same way a human neural network behaves and processes information.

Drs. Yann LeCun, Geoffrey Hinton and Yoshua Bengio were the beneficiaries of the prize for their work in a field that's been in constant development for more than 50 years. LeCun is currently chief AI scientist for Facebook, Hinton is based in the Google Brain research division, and Bengio is a professor at the University of Montreal and currently works alongside IBM.

Combined, their work on neural networking has changed the way AI tech is built and accelerated development in facial recognition, warehouse robots and autonomous cars.

The trio has been commended not just for their achievements, but their commitment to the field which saw a large dropoff between the mid-90s and mid-2000s. After combining to secure funding from the Canadian government to sponsor a research hub, they organised regular meetings, workshops and summer schools for their students.

It was in 2012 when their research led by Hinton beat the leading algorithm at the time for object recognition by 40% using their work on neural nets as a basis. That's when AI really start to take off and make people believe, according to LeCun.

It's a highly prestigious award that anyone in the field would love to receive, but who are the winners and what has led them to this great success?

Yann LeCun

Dr LeCun is a French AI specialist with a background in electrical engineering before making the switch to computer science, garnering various degrees from prestigious universities along the way.

LeCun now holds the director of AI research at Facebook and has published over 180 technical papers and book chapters on machine learning, computer vision, robotics, pattern recognition, neural networks, handwriting recognition, image compression, document understanding, image processing, VLSI design, and information theory.

His handwriting recognition technology is used by banks across the world and hisses image recognition model, convolutional network, is used by such companies as Facebook, Google and Microsoft for things such as image recognition and tagging and document recognition.

Geoffrey Hinton

Somewhat fittingly to his work with neural networks, British-Canadian Hinton's background lies in experimental psychology with a degree in the subject from King's College, Cambridge before gaining a PhD in artificial intelligence from the University of Edinburgh.

Some of Hinton's most notable work lies in the application of the backpropagation algorithm in neural networks which played a significant role in the way multi-layered neural networks perceived and understood data. He's also known for his work with Boltzmann machines which played a big role in the early understanding of neural networks.

Hinton now works at Google where he designs machine learning algorithms with an aim to discover a learning procedure that can be applied to large datasets in a way that mimics the human brain.

Yoshua Bengio

Bengio is regarded as one of the world's leading deep learning specialists who gained the most citations in academic literature in 2018. An engineering background provided him with the grounding to pursue further degrees in computer science.

According to his bio, his "ultimate goal is to understand the principles that lead to intelligence through learning". As well as work with IBM's deep learning department, Bengio has founded AI incubator in Montreal called Element AI which translates AI research into real-world business applications.

Nvidia and SAS have teamed up to connect businesses with AI technologies through Nvidia GPUs and CUDA-X AI acceleration libraries, combined with SAS's machine learning and data streaming technologies.

For example, SAS's Viya software can work with Nvidia's GPUs to help businesses identify objects and classify images, run speech-to-text applications and detect sentiment in communications, whether written or voice.

The tie-up is focused on the healthcare, life science, manufacturing and financial service sectors, helping those industries to connect better with their customers and ensure they deliver fault-free products to end-users.

"Analytics is the core of AI," said Laurie Miles, director of analytics at SAS UK and Ireland. "SAS is constantly looking at ways to improve its customers' analytics capabilities for successful AI deployments.

"This partnership with Nvidia combines our software expertise with their powerful GPUs, creating world-class engines for rapid, accurate AI operations. AI is opening up huge possibilities for organisations of all varieties, and SAS is at the forefront of innovation in this important field."

The partnership will also allow the firms to jointly develop new applications for their hardware and software that could be applied by tech companies in future. One example is equipping a drone with AI technology to monitor infrastructure and predict when machinery will need to go in to be serviced.

Because the analytics engine and the GPU are both implanted in the drone, the data can be analysed wherever the drone is, rather than sending the data back to another device to be analysed.

"Our collaboration with SAS will help enterprise customers extract the true value of AI for their company," said Ian Buck, vice president and general manager of Accelerated Computing at Nvidia. "With Nvidia technology, businesses will be able to accelerate their entire data science workflow to innovate, add new services and increase profitability."

SAS isn't the only company that Nvidia has partnered with in this area; the company has also announced this week that it has teamed up with major hardware vendors including Dell, HP and Lenovo to create purpose-built, high-powered workstations for data scientists also using Quadro RTX GPUs and CUDA-X AI libraries.

Researchers have trained a deep learning algorithm to detect signs of Alzheimer's Disease in patients on average six years before the condition is diagnosed by a human physician.

Californian-based scientists demonstrated that a neural network, once trained, was able to scan images of patients' brains and detect the presence of Alzheimer's on average 75.8 months prior to actual diagnosis.

The 20-strong team based their research on a modern diagnosis method, dubbed F-FDG PET (or fluorine 18 (18F) fluorodeoxyglucose positron emission tomography), in which a radioactive glucose dye is passed through the brain, and photographed.

Specialists then examine and interpret these images using the naked eye for signs of Alzheimer's, a precursor known as mild cognitive impairment (MCI), or other related conditions across the spectrum.

Despite seeming time-consuming, this method has led to quicker and earlier diagnoses, and more effective treatments.

But given this method is reliant on pattern recognition, researchers saw it as an opportunity to vastly improve its performance by deploying a self-training AI algorithm, publishing their findings in Radiology.

"There is wide recognition that deep learning may assist in addressing the increasing complexity and volume of imaging data, as well as the varying expertise of trained imaging physicians," the team wrote.

"The application of machine learning technology to complex patterns of findings, such as those found at functional PET imaging of the brain, is only beginning to be explored.

"We hypothesized that the deep learning algorithm could detect features or patterns that are not evident on standard clinical review of images and thereby improve the final diagnostic classification of individuals."

They set out to evaluate whether a deep learning algorithm could be trained to predict the final clinical diagnosis in patients who had undergone F-FDG PET, and how its success compared with current clinical standards.

From their study of 2,109 images from 1,002 patients who had already been diagnosed, they found their algorithm was able to detect Alzheimer's in images taken on average more than six years before diagnosis.

The algorithm performed better at recognising patients who would go on to have Alzheimer's than clinicians, as well as patients who would go on to develop neither Alzheimer's or its precursor MCI.

These findings are the latest in a series of studies and trials which show the potential power for AI to transform preventative healthcare and diagnosis.

In September the Francis Crick Institute revealed an AI learnt how to model and predict heart disease mortality rates in patients with a greater level of accuracy than trained doctors, or models created by experts.

Google's DeepMind AI project, meanwhile, reached an important milestone in the summer as its AI system was able to examine 3D images of the eye and diagnose sight-threatening conditions, as well as offer treatment advice, within seconds.

The algorithm, tested in conjunction with London-based Moorfields Eye Hospital, was able to recommend the best path of treatment for more than 50 eye diseases with 94% accuracy.

Despite bemoaning a handful of limiting factors, including a small sample size, the Californian researchers concluded that had developed a deep learning algorithm that can predict Alzheimer's "with high accuracy and robustness".

They added that with access to a much larger volume of data and opportunities to calibrate the model, the algorithm they developed could be integrated directly into the workflow of clinicians and serve as an essential support tool.

Researchers at MIT have created a neural network that can be used to spot the signs of depression in human speech.

In a paper being presented at the Interspeech Conference, the researchers detail a neural-network model that can be unleashed on raw text and audio data from interviews to discover speech patterns indicative of depression.

"The first hints we have that a person is happy, excited, sad, or has some serious cognitive condition, such as depression, is through their speech," says first author Tuka Alhanai, a researcher in the Computer Science and Artificial Intelligence Laboratory.

It is so advanced, the researchers say that given a new subject, it can accurately predict if the individual is depressed, without needing any other information about the questions and answers.

"If you want to deploy depression-detection models in a scalable way, you want to minimize the number of constraints you have on the data you're using. You want to deploy it in any regular conversation and have the model pick up, from the natural interaction, the state of the individual," said Alhanai.

It is hoped this method has the potential to be developed as a tool to detect the signs of depression in natural conversation, such as a mobile app that monitors a user's text and voice for mental distress and send alerts.

The researchers' model was trained and tested on a dataset of 142 interactions from audio, text, and video interviews of patients with mental-health issues and virtual agents controlled by humans.

Each subject was scored in terms of depression on a scale between 0 to 27, using a personal health questionnaire. Scores between 10 to 14 were considered moderate and those between 15 to 19 were considered depressed, while all others below that threshold were considered not depressed. Out of all the subjects in the dataset, 20% were labelled as depressed.

A key insight from the research was that during experiments, the model needed much more data to predict depression from audio than it did text. With text, the model accurately detects depression using an average of seven question-answer sequences. Whereas with audio, the model needed around 30 sequences.

"That implies that the patterns in words people use that are predictive of depression happen in shorter time span in text than in audio," Alhanai added.

Some of Google's datacentres are being cooled by AI, the search giant has revealed, with an algorithm controlling how fans and ventilation work to keep the equipment cool while using less power.

The project has been developed by Google's AI offshoot, DeepMind. It thinks the project could save millions of dollars of power savings and concurrently, reduce its carbon footprint.

DeepMind's datacentre cooling project uses the reinforcement learning type of AI, which works upon a system of trial and error to determine the most efficient way of cooling equipment, from a power consumption point of view.

Deep neural networks are fed information from the hundreds of sensors within data centre cooling systems every five minutes. This then predicts what different actions would do to its energy consumption, working out which would have the biggest impact on power. After safety verification, these commands are then sent to the data centre where they're verified and implemented again.

This type of reinforcement learning was also used by Google company DeepMind's AlphaGo robot that managed to beat professional Go players last year. Although it works completely autonomously, it can be taken over by a human if the action is thought to be too risky to the datacentre's operation or power consumption.

Previously, data centre operators were expected to manage the cooling system with AI recommendations, but it didn't implement the insights. After getting feedback from those managing the data centres on a day-to-day basis, Google decided to have the AI control the cooling system autonomously.

Google's AI control system is already providing energy savings of around 30 a month and has given the company the motivation to start testing it with other industrial applications to help tackle climate change "on an even grander scale."

The head of the Department for Digital, Culture, Media and Sport (DCMS) has said that new technologies like blockchain and artificial intelligence will be crucial in making Britain's economy a success following the UK's planned exit from the EU next year.

Speaking at the launch of a new cyber security innovation centre housed in Here East's Plexal tech campus in London's Olympic Park, secretary of state for digital Matt Hancock emphasised how important Britain's flourishing tech sector is to Britain's economy.

"Seizing this new technology can be the basis of our nation's success after Brexit, for the rest of the 21st century," he said. "What we need to do now to make that vision a reality. Because I profoundly believe that cutting edge technology, coupled with creative and artistic genius, is the fulcrum upon which our country will be built."

"Faster than ever before, the world we live in is being changed. How we earn our way around the world and how we build jobs and prosperity here at home, and we now have the chance to build on what we have done over the past decade and truly capitalise on this opportunity. Cutting-edge technology ultimately is the future of our economy."

Hancock pointed out that the UK's tech sector is currently booming; the last year saw tech investment double, and the area is growing at three times the rate of the rest of the economy. London is the top city in Europe for tech and globally, the UK is behind only China and the US in terms of its level of tech investment.

In particular, he said, Britain is renowned as an authority on AI, with numerous researchers and companies making great strides in the field.

The reason Britain is such a leading light in AI R&D, according to Hancock, is that it has strong cyber security, robust data protection principles and a good ethical standpoint, as well as firm links between academic institutions, businesses and governments.

"All of the great advances in the human condition have been led by improvements in knowledge and collective intelligence," Hancock stated. "This one is no different - except in that the intelligence is not just in the connection of human minds. Whether it's improving travel, making banking easier or helping people live longer, AI is already integral to our economy and our society."

One of the crucial issues the government is banking on AI to solve is healthcare. The government wants to diagnose 50,000 more cancer patients at early stages within the next 15 years using AI, opening up "properly safeguarded" NHS datasets to AI and med-tech companies in order to achieve this.

Hancock also announced that DCMS will be aided and advised by leading lights such as DeepMind co-founder Demis Hassabis, CognitionX co-founder Tabitha Goldstaub and Professor Dame Wendy Hall, who co-authored the government's AI review.

"Ultimately, AI will transform our lives like never before," he said. "And we want it to transform society for the better and for it to be designed and developed right here in the UK, because as we leave the EU, what better way to show that we are an open and outward looking nation than becoming the natural habitat for the world's most cutting edge technology?"

Microsoft CEO Satya Nadella yesterday highlighted the key role ethics must play in developing future AI technology.

Speaking at Microsoft's Leading Transformation in AI event in London, Nadella predicted AI will render enormous change in "every walk of life", but warned that developers, including those in his own company, would need to take their ethical responsibilities seriously.

"We're at that stage where the choices we make are grounded in the fact that technology development doesn't just happen - it happens because us humans make design choices," he said in a speech spanning a broad range of topics, including the Internet of Things (IoT) and quantum computing. "Those design choices need to be grounded in principles and ethics - and that's what's the best way to ensure the future we all want."

Speaking favourably of a recent House of Lords report into the need for an ethical code of practice in AI, he revealed that Microsoft has formed its own ethics committee, also stressing that the tech giant is working to ensure its technology reflected the high standards Nadella said Microsoft aspires to.

One of the key challenges in language, he said, "is the models that pick up language and learn from the corpus of human data available", lightly referencing Redmond's infamous Tay bot scandal - in which the company's Twitter-based chatbot, designed to emulate a teenage girl, began parroting racial slurs and conspiracy theories after being exploited by trolls.

"Unfortunately the corpus of human data is full of biases," he conceded, "so you need to invest in tooling that allows us to de-bias when you model language from the corpus of human data."

Joined by fellow Microsoft executives and customers from Unilever and Great Ormond Street Hospital (GOSH), Nadella outlined his vision for how AI could augment human capability, not replace it, mainly in language, audio and visual applications across a range of industries.

He cited the use of EmpowerMD at Pittsburgh Medical Centre, a software designed to cut down transcription time by intelligently transcribing doctor-patient conversations.

"The doctor is more efficient, and the patient is getting more of the doctor's attention and time, which is, in fact, the benefit of AI in the workspace," the CEO said. "And these design choices around we make in how we create AI and bring to market underline how we hope to make impact."

Unilever's CIO, Jane Moran, spoke at the event about how the use of machine learning has led to improvements in analysing and democratising data, before she demonstrated how its AI-powered chatbot, Unibot, is being used across the whole organisation to cut 90% of HR time spent on manual tasks like answering basic queries.

"Crucially, it frees up the HR professional's time to really focus on productive ways to support the staff in the business," she said, adding that the platform can be built out into other divisions in the company including legal, finance and manufacturing.

Meanwhile, Nadella outlined his vision for how AI is naturally integrated with cloud and edge computing, highlighting recent shifts in how the computing fabric itself is becoming distributed from the cloud to the edge - allowing developers to build multi-sense and multi-device experiences.

"What is perhaps at the core of the experiences we create, and the applications that we create, as well as the platform itself, is AI," said Nadella. "AI is the runtime that is going to shape all of what we do going forward in terms of the applications as well as the platform advances."

"We're reconceptualising the core products we build," he said, explaining the way the company's Office, Management and Security suites were rolled into Office 365, with the same principles applying to Dynamics 365, as well as its gaming arm, and LinkedIn.

Microsoft yesterday set out to open up its cloud platform to developers hoping to build AI applications, revealing a host of machine learning initiatives at its annual Build conference.

Redmond wants its Azure cloud to be the backbone of developers' AI innovations, with CEO Satya Nadella saying:"The era of the intelligent cloud and intelligent edge is upon us. These advancements create incredible developer opportunity and also come with a responsibility to ensure the technology we build is trusted and benefits all."

Project Kinect for Azure,is a package of sensors, including its next-generation depth camera, with onboard compute designed to allow local devices to benefit from AI capabilities.

Meanwhile, Microsoft's Speech Devices SDK aims to enable developers to build a variety of voice-enabled scenarios like drive-through ordering systems, in-car or in-home assistants, smart speakers, and other digital assistants.

The tech giant also previewed its Project Brainwave, an architecture fordeep neural net processingthat Nadella said "will make Azure the fastest cloud for AI". This is now available on Azure and for edge computing, and is fully integrated with Azure Machine Learning with support for Intel FPGA hardware and ResNet50-based neural networks.

The firm is also open sourcing the Azure IoT Edge Runtime, allowing customers to modify, debug and have more transparency and control over edge applications.

Azure IoT Edge also runs Custom Vision technology, a new service that Microsoft says enables devices such as drones and industrial equipment to work without cloud connectivity. This is the first Azure Cognitive Service to support edge deployment, with more coming to Azure IoT Edge over the next several months, according to the vendor.

"With over 30 cognitive APIs we enable scenarios such as text to speech, speech to text, and speech recognition, and our Cognitive Services are the only AI services that let you custom-train these AI capabilities across all your scenarios," said Nadella.

Also announced was a new SDK for Windows 10 PCs in partnership with drone company DJI. Using Azure cloud, the SDK brings real-time data transfer capabilities to nearly 700 million Windows 10 devices. As part of the commercial partnership, DJI and Microsoft will co-develop tools leveraging Azure IoT Edge and Microsoft's AI services to make drones available for agriculture, construction, public safety and other verticals.

Nadella also announced the company's new AI for Accessibility, a $25 million, five-year programme aimed at using AI to help more than one billion people around the world who have disabilities.

The programme consists of grants, technology investments and expertise, and will also incorporate AI for Accessibility innovations into Microsoft Cloud services. Microsoft said that the initiative is similar to its previous AI for Earth scheme.

Intelligent apps

Microsoft also handed developers the ability to introduce more customisation forMicrosoft 365 applications, so organisations can tailor them to their needs.

Developers whose businesses use the Microsoft 365 bundle of Office 365, Windows 10 and Enterprise Mobility and Security can now benefit from wider integrations in collaboration app Microsoft Teams, and even publish custom apps on the Teams app store.

There is also deeper SharePoint integration within Microsoft Teams to enable people to pin a SharePoint page directly into channels to promote deeper collaboration. Developers can use modern script-based frameworks like React within their projects to add more pieces that can be organised within SharePoint pages.

The vendor also revealed new Azure Machine Learning and JavaScript custom functions that let developers and organisations create their own additions to the Excel catalogue of formulas.

Picture: Satya Nadella/Credit: Microsoft Press Centre

The Mayor of London and AI research centre CognitionX plan to conduct a study into the potential of AI technology and how it's already being used by companies based in the capital to solve problems.

It will highlight the support businesses need to grow the industry sector, including developing and deploying solutions, fnding the right talent to fill skills gaps and closing barriers to adoption, such as security concerns.

"This report will help to uncover the opportunities to unlock innovation and investment in London, in order to maximise the economic impact of AI on the city," Mayor of London Sadiq Khan said. "It will also identify the challenges we face in positioning the capital as the best place to start, grow or relocate an AI business. Artificial intelligence has the potential to transform almost every industry across the capital."

The study will call upon businesses working in the AI space to explain how they're developing technologies to solve problems and make life better for Londoners.

"London is in a strong position in the data economy and is already home to innovative, fast-growing companies like Deepmind, CityMapper and Satalia not to mention the kind of work being done to improve public services, such as the data-driven approach to understanding rent arrears emerging through Hackney Council's partnership with Pivigo," Khan continued.

Businesses that want to get involved in the research can fill in a form on CognitionX's website, offering information about how they've deployed AI and how they anticipate using it in future. The information will be collated in the report and distributed at AI festival CogX on the 11th and 12th of June as well as appearing on the Mayor's Tech Map London platform.

"We all need to understand what AI is, what it isn't and how we can benefit," Charlie Muirhead, founder and CEO of CognitionX said. "At CognitionX we are on a mission to democratise access to information on AI and this report is an important part of CognitionX's commitment to being the most trusted source of advice on all things AI.

"All successful organisations will need to take advantage of artificial intelligence, but it is a complex, fragmented and ever-changing domain and we want to help ensure a responsible transition to an AI-driven society."

Nvidia taught MPs, peers and government staff about how to build neural networks last week, in an effort to raise awareness about how the technology can help local communities.

Those in attendance were also given an overview of how AI is currently being used in the UK's infrastructure, such as transport, healthcare and building smart cities.

Stuart Wilson, AI and supercomputing director at Nvidia, said: "Artificial intelligence represents the biggest technological and economic shift in our lifetime. It is of national importance that policy makers understand the core components, capabilities and limitations surrounding the modern AI boom."

Although Nvidia, best known for its graphics processing units, doesn't expect MPs to start building their own neural networks, the session was designed to educate staff so they can make better decisions around AI and respond to constituent questions about the technology.

The workshop was organised by Big Innovation Centre in collaboration with Nvidia, with presentations from UCL Hospital Neurology Department, Bristol-based autonomous vehicle specialist FiveAI and Milton Keynes City Council about how they are using AI to improve the public's lives.

"Government and policy makers are not technology experts," Professor Birgitte Andersen, CEO of Big Innovation Centre said. "Practical, accessible examples of AI in action provides a real-time understanding of the impact and implications of AI for our businesses and society. More informed decision-making allows the UK to get AI-ready as a leading global innovator.

"This workshop is the latest step in ensuring Parliamentarians and other key stakeholders have the knowledge and tools they need to shape policies, regulations, and budgets in our AI future."

Last week the government was accused of having made a "massive strategic error" on AI funding, with experts raising concerns that Britain was falling significantly behind its western rivals in investment, and that Brexit would likely cut off essential cash flows from the EU.

Image: Shutterstock

A consortium of tech companies including AWS, Nvidia and Scan Computers is setting out to introduce children to the concept of AI technologies.

The firms have teamed up with Beverly Clarke, a Computing At School (CAS) Master Teacher, to help Year 9 pupils learn about the terminology used in the world of AI. It will also seek to help children understand how AI is being used in the real world, with guidance to help them create their own image recognition neural network.

"AI is already part of our everyday lives, and by the time today's 13-year-olds are entering the workforce, it will have a significant impact on the kinds of jobs available to them," Clarke said.

"The World Economic Forum estimates that, by 2025, 90 percent of jobs will require digital skills, and that 65 percent of children entering primary school today will work in jobs that don't currently exist. It's critical that we introduce pupils to core AI concepts so they're equipped to thrive in this environment."

The programme has been developed to adhere with the Key Stage 3 requirements, with lesson plans, worksheets and activities to help teachers educate their pupils.

"Education underpins our ability to embrace and exploit the promise of AI," said James McClung, Higher Education and Research Business Development Manager at Nvidia. "Young people's lives will be infused with AI, from their homes and transportation to the workplace. They need to understand this technology so they can think critically about it and consider what career options it might open up for them."

The programme will initially be tested in six schools and if the pilot is successful, secondary schools across the country will be able to sign up to it.

Of all the government agencies to make a major announcement, NASA's has got to be the most exciting. The space agency doesn't make banal announcements. If it warrants a press release, you can be damn sure it's worth our attention.

During the most recent NASA announcement, the space agency revealed it has spotted an eighth planet circling Kepler-90, a Sun-like star 2,545 light years from Earth.

The planet, dubbed Kepler-90i - was discovered in data from NASA's Kepler Space Telescope after being analysed by Google's AI. It is a "sizzling hot, rocky planet" that orbits Kepler-90 once every 14.4 days.

Google trained its AI to learn how to identify planets by searching Kepler data for instances where the telescope recorded signals from planets beyond our solar system, known as exoplanets.

The discovery of this eighth planet means our solar system is now tied for most number of planets around a single star.

"Just as we expected, there are exciting discoveries lurking in our archived Kepler data, waiting for the right tool or technology to unearth them," said Paul Hertz, director of NASA's Astrophysics Division in Washington. "This finding shows that our data will be a treasure trove available to innovative researchers for years to come."

The AI was trained by Christopher Shallue and Andrew Vanderburg to look for signals associated with exoplanets. This include changes in light readings recorded by Kepler as the exoplanets orbit their Sun. Tiny dips in brightness captured when a planet passes in front of, or transited, a star can be used to identify the planet, but also determine its orbit and size.

Inspired by the way neurons connect in the human brain, Google's artificial "neural network" sifted through Kepler data and found weak transit signals from a previously-missed eighth planet orbiting Kepler-90, in the constellation Draco.

This was no mean feat, however. Kepler's four-year dataset consists of 35,000 possible planetary signals.

The AI was trained using a set of 15,000 previously-vetted signals from the Kepler exoplanet catalogue, and was able to correctly identify true planets and false positives 96% of the time.

Then, with the neural network having "learned" to detect signals, the researchers told their model to search for weaker signals in 670 star systems that already had multiple known planets. Their assumption was that multiple-planet systems would be the best places to look for more exoplanets.

Their research is published in The Astronomical Journal. Shallue and Vanderburg now plan to apply their neural network to Kepler's full set of more than 150,000 stars.

London-based Chattermill has closed a seed round of 600,000, gaining funds that should help it boost its development of machine learning technology.

The deep learning startup applies artificial neural networks to customer feedback that learn from a company's data to help it make better decisions. It attempts to measure customer feelings about the design of the company's app, the speed of delivery and how customers feel about customer care agents.

The company was co-founded by Mikhail Dubov and Dmitry Isupov in 2015 who met at Entrepreneur First, a London-based company builder and startup accelerator that introduces potential co-founders to each other in order to form a team.

The total amount Chattermill has raised is 935,000 as the startup had an earlier round of funding in 2016 from the same group of investors. The funding will be used to invest in technology and continue growing the team with new hires including data scientists, engineers and business developers.

Co-founders Dmitry Isupov and Mikhail Dubov

Its product easily integrates with a number of tools, such as SurveyMonkey, Zendesk, TypeForm or Salesforce, and aggregates every piece of feedback to produce a thorough analysis of customer experience. The startup is currently working with customers across sectors that include, fintech, e-commerce, travel and gaming.

Chattermill has been backed by Entrepreneur First, Avonmore Developments and angel investors including Jeff Kelisky, CEO of Seedrs.

"We're thrilled to have the ongoing support of such a great list of investors," said Mikhail Dubov, the CEO and co-founder. "We've been lucky enough to help some of the world's most customer centric businesses see genuine value by understanding their users at scale. Our platform not only challenges their assumptions, but gives them incredibly detailed insight in real-time, at a fraction of the cost of traditional customer experience research."

Chattermill has a number of fast-growth businesses as clients already including Transferwise, Deliveroo, HelloFresh Just Eat, and TSB.

"Chattermill enables our teams to take customer insights deeper than ever before and focus on the key factors that make a difference to our users and drive our growth," said Nilan Peris, VP of growth at Transferwise. "We have more than a hundred people across the business using Chattermill to ensure Transferwise is always at the cutting edge when it comes to customer experience."

LinkedIn co-founder Reid Hoffman invested in Entrepreneur First in September. His venture capital firm Greylock Partners led a $12.4 million funding round and was joined by Lakestar Capital and the founders of Deepmind.

"We saw the opportunity in Chattermill from the very beginning and are happy to continue to back Mikhail, Dmitry and their growing team," said Matt Clifford, CEO and co-founder of Entrepreneur First. "They are a fantastic example of how machine learning and AI can dramatically impact product development by listening to what the customer really needs. We're all very excited to see what happens in the coming year for Chattermill and see what other learnings they can provide their customers."

Chattermill joins a number of UK AI startups that have raised funding including Graphcore, who raised $50 million in November, as well self-driving car startup FiveAI who raised 26.8 million in September.

Image sources: Chattermill

Google has announced the AIY Vision Kit, opening up a new way for Raspberry PI users to access its AI tools. The kit comprises a low-power Intel Movidius MA2450 circuit board and computer vision software that can be tagged onto an existing Raspberry Pi computer and camera modules.

The circuit board provides the ability to run neural networks on the device rather than having to add on an external source for processing the data or use a cloud processing system.

The software includes access to three TensorFlow-based neural network models - one based on MobileNets to recognise objects, one to recognise faces and human facial expressions, while the third can identify the difference between a human, a cat and a dog.

"We've also included a tool to compile models for Vision Kit, so you can train and retrain models with TensorFlow on your workstation or any cloud service," Google's Billy Rutledge, its director of AIY Projects said.

"We also provide a Python API that gives you the ability to change the RGB button colors, adjust the piezo element sounds and access the four GPIO pins."

Google said its AIY Vision Kit could be used for a whole range of applications, including identifying plants and the individual species, monitor where your dog is in your home, see when your car left the driveway and recognise emotion - for example, whether your guests like your home decor.

This is the second product released in the AIY line by Google. Its AIY Voice Kit was released in May and enables Raspberry Pi developers to build a standalone voice recognition system using Google Assistant, or users can add their own voice recognition and natural language processing to projects.

Image courtesy of Google

Understanding foreign languages has always been a barrier for individuals and businesses seeking to expand in other countries. Language learning and translation has somewhat aided this and with the recent technological advancements in artificial neural networks this may become even easier.

In simple terms, artificial neural networks normally shortened to "neural networks" are a type of artificial intelligence (AI) that mimics the biological neural networks seen in animals.

Erol Gelenbe, a professor at Imperial College London's department of electrical and electronic engineering, is one of the leading researchers in the field, whose interest in artificial neural networks (here on shortened to "neural networks") developed from his earlier work on anatomy. He started by trying to build mathematical models of parts of human and animal brains, then graduated to using neural networks to route data traffic across the internet and other large networks.

Gelenbe says translation has three different aspects, whether carried out by a machine or a human being. The first is word to word translations, which can be accelerated or simplified using neural networks and other fast algorithms. The second is mapping the syntax, which means the neural network will have to "understand" the nuances of grammar in both languages. The third is using context to translate, which is extremely important as it directly affects which words are chosen.

Gelenbe uses English and German as an example: "Neural networks can be used for each of these steps as a way to store and match patterns, for example matching school' with schule', matching to' with nach', or learning and matching the grammatical structures".

A robotic rosetta stone?

Google and Microsoft both introduced neural machine translation back in November 2016. It differs from the previous large-scale statistical machine translation, as it translates whole sentences at a time instead of just one or two words at a time. In a blog post, Google explained how the sentence is translated in its broader context and is then rearranged and adjusted "to be more like a human speaking with proper grammar". This makes it easier to translate larger bodies of text as they are taken sentence by sentence, so paragraphs and articles will be translated with fewer errors or instances of miscomprehension. Microsoft has a useful tool to highlight the difference between neural networks and statistical machine translation, which shows how neural translation sounds much more natural. And the best part? Over time neural networks learn to create better and more natural translation.

But neural network-powered translation isn't all about completely new innovations - it also builds technologies being used in other domains, such as LSTM.

LSTMs support machine learning and can learn from experience, depending on how they are applied. Since 2015, Google's speech recognition on smartphones has been based on self-learning long short-term memory (LSTM) recurrent neural networks (RNNs) and the technology has been extended to other products, including Google Translate.

Jrgen Schmidhuber, professor and co-director of the Swiss Dalle Molle Institute for Artificial Intelligence and president of NNAISENSE, who developed LSTM-RNN technology, predicts that in the future these systems will enable "end-to-end video-based speech recognition and translation including lip-reading and face animation".

"For example, suppose you are in a video chat with your colleague in China. You speak English, he speaks Chinese. But to him it will seem as if you speak Chinese, because your intonation and the lip movements in the video will be automatically adjusted such that you not only sound like someone who speaks Chinese, but also look like it. And vice versa," Schmidhuber explains.

Neural knowledge

Towards the end of April 2017, an update was released for Google Translate allowing it to decipher between English and nine Indian languages using neural machine translation technology. Previously, Translate only supported translations between English, French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish. This was an essential step, as for multilingual countries like India, which has 23 official languages, translation is often required for domestic communication, not just international.

Google highlights that: "This new technique improves the quality of translation more in a single jump than we've seen in the last ten years combined". Yet adding more languages also had an unexpected benefit: that neural technology speaks a language better when it learns several at a time, similar to how it's easier for humans to learn a language when they know a related one. This means if there isn't a lot of sample data for one language, such as Bengali, but there is a lot for another, like Hindi, then the translation is able to use one to intelligently fill the gaps in the other.

Google Brain, the company's dedicated deep learning and AI research project, recently announced researchers there are using neural networks to aid language translation from speech to text. A user speaks the language they want translated and it's then written in a different language straight away eliminating the need for a transcription of the text in the original language.

Translating speech directly to text can be useful in a number of ways. One example is Robot Lawyer's Do Not Pay, which was originally built to help people work out if they had to pay a parking fine or not, but now also helps people apply for refugee status in countries where they will typically not speak the native language. Another example is Babylon, an AI-powered health app that helps users diagnose themselves through a series of questions. Opening it up to other languages increases the number of people it can help.

Improving translation

According to Rick Rashid, founder of Microsoft Research, in the 10 years between 2000 and 2009 there was no change to word error rate in automatic speech recognition (ASR) software that transcribes speech into text.

But, thanks to the development and implementation of deep-learning neural networks, researchers at Microsoft, the word error rate in translations was improved drastically.

"In 2012 I was able to stand up on stage in Tianjin, China and have my own voice simultaneously translated from English to Chinese live on stage," Rashid says. "This was a testament to the huge improvement in word error rates and to the translation technology we put in place."

This technological jump could aid the business world, where organisations span multiple countries and continents and where clients or colleagues often don't speak the same language. In particular, it can help SMBs grow faster and expand into areas of the world where the language barrier would otherwise present a significant obstacle.

It may have a huge impact on cross-industry learning too. Communication across borders would be vastly improved, and the sharing of information and data would be made more efficient as a result. More business models can be adapted and put to use in different industries to foster ground breaking change.

There's still much more to be explored with neural networks and Gelenbe still thinks the most exciting neural network discoveries await us.

"To understand how our brain does very complicated things very quickly and so efficiently, is still before us - the future will be more exciting than the past."

Image sources: Header: Bigstock; First image: Google; Graph: Rick Rashid