Imagination’s Question of the Month: Is AI overhyped?

Welcome to the first of new series of blog posts where on a monthly basis we will be exploring some of the big issues facing the industry right now. We will do so by asking a question to experts in the industry and also the expertise within our company. The first question is simply: 'is AI overhyped?'

Imagination recently entered the world of AI-related industry with the recent announcement of its neural network accelerator (NNA), the PowerVR Series2NX. This offers hardware acceleration of neural networks at unprecedented levels of performance and will enable upcoming embedded and mobile devices to take advantage of neural network powered AI applications.

But has too much stock been placed in AI as a whole? As you will no doubt be aware, AI is taking over headlines everywhere with the promise to reorganise, revitalise and revolutionise our world. Of course, there are arguments for and against and many prominent, public figures have expressed their concerns over its potential impact. Just this week Hilary Clinton stated that America was, 'totally unprepared' for the impact of AI, while the likes of Elon Musk and Stephen Hawking have both expressed their deep fears over where it could leave the human race.

So what does Imagination think? Rather than presenting a single unified answer in this post we get the personal views of five members of our staff.

Question: Is AI overhyped?

Bryce Johnson, Director Automotive Segment Marketing
Is AI overhyped? - NO

We are only at the beginning of the deployment curve of the application of AI to everyday things and one can only dream of a future where AI is the fundamental technology to address massive problems that we are facing in the world today. There is so much value that can be extracted from AI that will enable our lives to be richer and allow us to focus on the more important things in life.

Let's consider the move to the autonomous car. Without computer vision (a subset of AI), the car would not be able to make sense of the world around it and enable the car to be controlled in complex environments. Without AI in the Cloud to perform the 'training' of such vision systems the AI would be prone to error. Without machine learning (another AI subset), the car would not be able to learn from the new experiences it comes across as it is driving around.

This is just one example where AI will make a huge difference to our lives. It will reduce accidents (91% of all accidents are caused by human error), reduce pollution by improving traffic flow, and reduce congestion with advanced car share and car ownership schemes.

The automotive space is one where the promise of AI has still not been realised to anywhere close to its full potential.

Dave Bartlett, Senior Software Engineering Manager, Vision and AI Software
Is AI overhyped? - YES

This question can be considered from a number of different viewpoints and over a variety of timeframes. At the general, public level, AI is massively over-hyped in terms of the capability of the technology and the impact on society. Every day I read stories in mainstream media trying to explain to the general public the impact of AI. Recently, Radio 4's 'PM' program, a show that is regularly listened to by over seven million people, dedicated a slot every day for a week to explore the impact of AI on various sectors of the economy.

Futurologists regularly predict that within 10 years AI will start to have a major disruption on society and the classic example given is the self-driving car. And what do we learn from history - that futurologists are always wrong!

While it's true that the technology is advancing quickly, what we have today is very basic cannot be considered AI and is better described as machine learning algorithms. Let's consider today's state-of-the-art classification networks.

These networks are shown a picture and classify it (like the not hot-dog detector). These networks are then trained using thousands of images and are often claimed to be better than humans at the task for which they have been trained. However, research into the limitations of these networks has generated some very interesting results. A team at MIT were able to 3D print a turtle that was classified as a rifle!

This was achieved by manipulating the pattern on the turtle. To a human, the 3D model was undoubtedly a turtle but it fooled the neural network. Equally, other researchers have fooled classification networks by changing a single pixel in an image. These networks are very easy to fool because they are primitive and there needs to be a massive improvement in their robustness before they can be used for safety-critical applications. The car industry has talked about self-driving cars by 2020, but this will not be possible with today's technology.

To summarise, technology bubbles are common, futurologists are always wrong and the impact on society from technological advances is gradual. AI technology will rapidly be deployed in AR, VR, human interaction and many other areas and it will drive revenue and grow in the tech sector. However, to match today's hype, massive improvements in the technology will be required.

Russell James, Vice President of Vision & AI
Is AI overhyped? - NO

The surprise has been that AI has not followed the hype curve in the traditional sense. The technology was hyped but has gone straight into production - to reality - without entering the 'trough of disillusionment'. This is rare and has fuelled a more rapid adoption of this technology and what it can offer. I believe that AI will be one of the driving technologies of 2018 with adoption into consumer, commercial, robotics, IoT and datacentre, and longer term into automotive. I'd say that pretty much every corner of the semiconductor industry will be impacted by AI technologies.

Francisco Socal, Product Manager, PowerVR Vision & AI Business Operations
Is AI overhyped? - YES

This is a very relevant question today. Overall, I would say yes, there has been a significant amount of hype. This is due to is a mismatch between current expectations and enthusiasm for AI and what it can actually deliver today.

What many people might not be aware but this is not the first wave of AI tech. It's actually the third or arguably even the fourth. The first one was in the 1950s when the link between human intelligence and machines was first observed. However, the technology and performance were not in place and the outcome, in reality, turned out to be disappointing. Decades later, with the arrival of high-performance GPUs and compute capable software it could be said the that the 'AI winter' was definitively over.

As with any new technology though, it inevitably goes through the classic 'hype cycle' and in my view, I would say it's overhyped. What we have today is still far from general artificial intelligence that many have said with displace humans by eliminating jobs. In this paper, even the most enthusiastic experts have an expectation that it will take at least 30 years before AI can perform high-level machine intelligence (HLMI) when unaided machines will be able to accomplish tasks better and more cheaply than human workers. Others say it will take at least another 75 years!

We are only at the start of this journey. There are a collection of related technologies and disciplines related to AI and it is important to discern what they are. For example, 'deep learning', of which convolutional neural networks are examples, are in the spotlight right now, attracting lots of interest and investment. Specifically with regard to these, while they are hyped I would say not overly so. The reason is so that they can deliver value now by solving real-world applications with significant improvements over previous methods. These will find their way across a multitude of applications across very diverse industries.

They are starting to come under scrutiny, however, and have been shown that they can be fooled, which could cause them to fall into the trough of disillusionment. Examples are how an autonomous car can be confused simply by attaching stickers to road signs or my all-time favourite, the neural network that was tricked into thinking a 3D-printed turtle was a rifle. (Spoiler: it wasn't).

However, there is more to deep learning than AI. For example, speech recognition, speech synthesis and natural language processing rely heavily on neural networks and are finding their way into very practical applications in mass usage consumer products, delivering results with very high accuracy.In that sense, expectations are matching the actual benefits, so there's no over-hype at all.

On the other hand, other AI techniques are still emerging such as reinforcement learning and neuromorphic processing and while we are yet to see the hype around these, I would expect these to see significant growth and take the spotlight from CNNs.

David Harold, VP Marketing and Communications
Is AI overhyped? - NO

I think we have to look at AI the way we look at spaceflight. Today the benefits are meaningful but relatively small - better materials, better engineering - but the potential is that we unlock all the resources of the solar system. AI is similar. Today it can beat Go, tomorrow it might help us allocate resources so we can live in a post-scarcity world. It's interesting that Stephen Hawking, a man not known for hyperbole, regards AI as one of the most significant (and potentially threatening things) to happen to humankind. He takes a long view, somewhat pessimistically. I prefer the approach of Iain M. Banks who envisaged a world in which we have handed over responsibility for our wellbeing to immensely more rational AIs, freeing us to live complete and fair lives.

I agree with what Bank's said: 'I think AIs might be the saving of us,' but I also agree with his idea that 'ideally you want strength-in-depth in your future society, with a robustly reliable step-down process in place that lets you fall back to the dumb-but-fast supercomputers and on down to standard human capabilities-without suffering too much in the process-just in case the Minds, or whatever, do suddenly decide they're bored with the whole thing and disappear up their own collective fundament, possibly taking any other potentially sapient ware with them.'

That's the future of Frank Herbert's Dune of course, where the AI-era has been and gone.

We're a long way off an AI-era of course, but it's starting to look like AI could be as important to this century as the automobile was to the 20th.

To keep up to date with Imagination Technologies you can follow us on Twitter at @ImaginationTech and on LinkedIn, Facebook and Google+.