Recent Must Read AI Books

End of last year I finally managed to start reading more often again. It also helps when you read engaging books since you can’t put them down. From the selection I read, three books stand out. Three AI related books, where they are close on the topics they discuss, especially two of them. By reading order, the books are:

Rebooting AI: Building Artificial Intelligence We Can Trust

This is a book where Gary Marcus and Ernest Davis do a good job of describing where the AI field actually is, what are the strengths and weaknesses of current technologies (specifically Deep Learning), where they think the area is going wrong, and possible paths to move forward. The main point of the book is that, in spite of the recent advances and specific practical breakthroughs, these are still narrow and fuel a lot of the current hype around AI. The book goes through an exhaustive list of examples where it shows how Deep Learning based solutions fall into these issues, and most likely are not the path for general AI. Although the book is very good in showing the weaknesses of current approaches, it falls short on presenting new directions and it avoids or goes light on the issues of past approaches, especially the ones coming from symbolical AI. The authors advocate the need of hybrid systems, but in many parts it feels it’s mostly about bringing symbols to neural nets, which in itself is also not enough for the kind of AI they want to evolve. Overall, it’s a good book that should be read since it makes you think more critically about the current state of the field, no matter which side you’re coming from.

Artificial Intelligence: A Guide for Thinking Humans

Melanie Mitchell book is very close to the topics that Rebooting AI deals with. However, it’s a superior book in all its forms. Overall, it starts with the generic question “How intelligent really are the current best AI programs today?” and from here it starts a quest describing the field of AI, the past, present and also what the future could hold. The writing, pace and the depth of each topic is excellent. Everything is presented in very balanced and fair way. She also discusses the disconnection between AI research/applications and the media hype. And all this is done with a great sense of honesty and accessibility. It also intervenes with the stories of the people and science. If GEB was a book that defined a reading generation, this one has the potencial to define a new younger generation too. This book is definitely the must read AI book of today, no matter what is your knowledge, position, etc.

Human Compatible: Artificial Intelligence and the Problem of Control

Finally, Stuart Russell latest book is mostly about the topic of AI safety, especially with the danger of having achieving AGI which can turn against its creators. and how to deal with it. The main point of the book is that we need to change the way we develop AI. Especially if we want AI to remain beneficial to us. The proposal is to use  Inverse Reinforcement Learning which means having AI learn our preferences and goals by observing us, instead of optimizing for a goal. Stuart Russel does a great job of describing all the issues, theory, proposals and then his proposal to use IRL. This somehow connects also to the two previous books in the sense we need to build different AI systems, consider alternatives (e.g., in this book he makes a nice case for evolutionary systems), but most importantly, design with safety in mind. This is also a great read!

To conclude, these are three great books that I believe everyone with an interest in AI should read. Overall, they provide a good overview the past and present if the AI field but, more important, how to think and consider different approaches. This will be a requirement to truly advance the field!