Book Review: Machines Like Us, The Difficulty Of Coding AI For Common Sense

When this column reviews books, it’s usually about one that directly links technology to a horizontal or vertical business sector. However, sometimes I get sent one with a wider picture. “Machines Like Us: Toward AI With Common Sense,” Ronald J Brachman and Hector J. Levesque, The MIT Press, is one such and was a very interesting read. It’s a high level view of what artificial general intelligence (AGI) is so difficult.

In my opinion, this book is good for anyone wanting to get a broad understanding of the challenge of AGI. It does have some code and pseudo-code segments but they are simple and clear, used to support text points. While the book will serve IT management well in helping to understand how to question vendors pitching artificial intelligence (AI) solutions as broader than they are, it’s also good for anyone wondering about the spread of AI in just about all aspects of our lives.

Computer software, of all types, has become very good at doing what we know and then pointing out exceptions. However, it’s common sense that helps us deal with new situations, the exceptions, without anyone having to rewire our brains. To me, common sense is another way of saying exception handling. That is still needed for AI systems to really work well even in restricted spaces, but the lack of it is a good example of a key reason automated vehicles still aren’t here.

I’ve always loved a much shorter version of a Voltaire attributed quote that simple says “common sense isn’t.” After a brief discussion of common sense, the authors have a nice introduction to the five main areas of AI. I could quibble with some, but for people trying to understand the point of the book, it’s an excellent overview of what AI is at this point in time.

After that, the book turns to how to represent knowledge and common sense. Given my background, I was pleased to see then talk about rules-based systems, even using the much out of favor noun that is expert systems. They describe both the power and the failings of those for trying to replicate common sense. There is also a discussion about deep learning, but it’s kept brief without the detail that many books use.

Over the years, a number of articles in this column and elsewhere have pointed out that some of the most interesting AI systems use deep learning, expert systems and procedural code in concert to understand data and to deliver intelligible information. Any system to successfully use common sense to handle exceptions will use a blend of these and probably something that hasn’t yet been invented.

Chapter nine is a key chapter. After the overview of common sense, AI systems, and representational constructs, it focuses on how to implement a common sense system. Towards the end of the chapter, just as it looks like there needs to be a blend of technical tools and techniques, the book points out that both top-down and bottom-up approaches have advantages and will also need to be blended.

Too many people claim one of two extremes. Either AGI is hard so it will never be solved, or their own tools are so special those will somehow be the answer. I know this will shock the audience, but our brains are complex and we still don’t understand them. However, they do make some people smart. By continuing to investigate multiple avenues we extend our knowledge.

Machines Like Us, for those who want to better understand a key part of intelligence we still haven’t been able to clarify and build for artificial intelligence, is an excellent overview.