Skip to main content

How Smart Machines Think - Sean Gerrish ****

While it will become apparent I think this book should have been titled 'How Dumb Machines Think', it was a remarkably enjoyable insight into how the well publicised AI successes - self-driving cars, image and face recognition, IBM's Jeopardy! playing Watson, along with  game playing AIs in chess, Go and Atari and StarCraft, perform their dark arts.

There's no actual programming presented here, so no need for non-programmers to panic, though there is some quite detailed discussion of how the software architectures are structured and how the different components - for example neural networks - do their job, but it isn't anything too scary if you take it slowly.

One thing that comes across very strongly, despite the AI types' insistence that their programs are of general use, is how very specifically tailored programs like the AlphaGo software that beat champions at the game Go, and the Watson computer that won at the US TV quiz show Jeopardy! were - incredibly finely designed to meet use and that use only.

The reason I make the remark about dumb machines is that what doesn't come across sufficiently in Sean Gerrish's book is that, because these programs are not in any sense intelligent, when they get things wrong, they often get things dramatically wrong. So some of Watson's answers on Jeopardy! did not make any sense at all. Similarly, image recognition software can be fooled by apparently abstract patterns that happen to have the right components to appear to be a distinguishable object. And when you bear in mind we're suggesting putting this kind of software in charge of cars that 'getting it dramatically wrong' bit is more than a little unnerving.

There was, though, a great section on the development of self-driving cars, from the original feeble attempts, where all the competitors in a race failed before completing 10 percent of the course, through to more recent and more successful versions that can handle basic traffic scenarios - though it would have been nice if Gerrish had gone beyond the old prize challenges to describe what the latest Google and Uber vehicles do. (It may be that their approaches are too proprietary.)

However, what was missing was any serious assessment of the big problems still faced. There have been two excellent books recently on the huge holes in AI that practitioners rarely admit to - The AI Delusion and Common Sense, The Turing Test and the Quest for Real AI - Gerrish would have produced an even better book if he could have addressed the concerns that these books raise.

Even so, in How Smart Machines Think we have a hugely informative and very readable book for anyone with an interest in finding out just what the much-trumpeted AI systems really do, and what lies beneath the hype.

Hardback:  

Kindle:  

Review by Brian Clegg

Comments

Popular posts from this blog

Beyond Weird - Philip Ball *****

It would be easy to think 'Surely we don't need another book on quantum physics.' There are loads of them. Anyone should be happy with The Quantum Age on applications and the basics, Cracking Quantum Physics for an illustrated introduction or In Search of Schrödinger's Cat for classic history of science coverage. Don't be fooled, though - because in Beyond Weird, Philip Ball has done something rare in my experience until Quantum Sense and Nonsense came along. It makes an attempt not to describe quantum physics, but to explain why it is the way it is.

Historically this has rarely happened. It's true that physicists have come up with various interpretations of quantum physics, but these are designed as technical mechanisms to bridge the gap between theory and the world as we see it, rather than explanations that would make sense to the ordinary reader.

Ball does not ignore the interpretations, though he clearly isn't happy with any of them. He seems to come clo…

The AI Delusion - Gary Smith *****

This is a very important little book ('little' isn't derogatory - it's just quite short and in a small format) - it gets to the heart of the problem with applying artificial intelligence techniques to large amounts of data and thinking that somehow this will result in wisdom.

Gary Smith as an economics professor who teaches statistics, understands numbers and, despite being a self-confessed computer addict, is well aware of the limitations of computer algorithms and big data. What he makes clear here is that we forget at our peril that computers do not understand the data that they process, and as a result are very susceptible to GIGO - garbage in, garbage out. Yet we are increasingly dependent on computer-made decisions coming out of black box algorithms which mine vast quantities of data to find correlations and use these to make predictions. What's wrong with this? We don't know how the algorithms are making their predictions - and the algorithms don't kn…

Five Photons - James Geach ****

It is generally acknowledged that Stephen Hawking's A Brief History of Time is one of the most common books to be bought but not read beyond the first few pages. If you are the kind of popular science reader who found Hawking hard going, you can stop now - Five Photons is not for you. If, on the other hand, you found A Brief History of Time a piece of cake and wished you could get into more depth without resorting to heavy mathematics or a tedious textbook style, Five Photons could be just up your street.

Astrophysicist James Geach starts of fairly gently with a chapter on the nature of light that mostly sets aside quantum physics, leading up to the observation that light is our vehicle for for stripping back the history of the universe to its earliest times (or, at least, the point where the universe became transparent). From here on, the five photons of the title take us on different journeys, from the oldest surviving light of the cosmic microwave background radiation to that fr…