Natural General Intelligence: How understanding the brain can help us build AI

by Christopher Summerfield

Since the time of Turing, computer scientists have dreamed of building artificial general intelligence (AGI) - a system that can think, learn and act as humans do. Over recent years, the remarkable pace of progress in machine learning research has reawakened discussions about AGI. But what would a generally intelligent agent be able to do? What algorithms, architectures, or cognitive functions would it need? To answer these questions, we turn to the study of natural intelligence. Humans (and many other animals) have evolved precisely the sorts of generality of function that AI researchers see as the defining hallmark of intelligence. The fields of cognitive science and neuroscience have provided us with a language for describing the ingredients of natural intelligence in terms of computational mechanisms and cognitive functions and studied their implementation in neural circuits.

Natural General Intelligence describes the algorithms and architectures that are driving progress in AI research in this language, by comparing current AI systems and biological brains side by side. In doing so, it addresses deep conceptual issues concerning how perceptual, memory and control systems work, and discusses the language in which we think and the structure of our knowledge. It also grapples with longstanding controversies about the nature of intelligence, and whether AI researchers should look to biology for inspiration. Ultimately, Summerfield aims to provide a bridge between the theories of those who study biological brains and the practice of those who are seeking to build artificial brains.

Genres: Artificial Intelligence, Neuroscience, Technology, Nonfiction, Psychology

352 pages, Hardcover

Published March 15, 2023

Reviews

[Tero Parviainen · Rating 5 out of 5]

It's an absolute treat to have machine learning and the mind sciences used as lenses to examine each other in such a methodical but accessible way. You end up with a greatly enriched understanding of both.

[bimri · Rating 5 out of 5]

Intellect in nature is a precarious thing to grasp. Hence the many cycles AI research has had to go through to even bring forth products as AlphaFold, MuZero, chatGPT or even Stable-diffusion models.

Summerfield has done such a gorgeous and brilliant job at writing this book. Gorgeous for the creativity exhibited to write a book that was so precisely elegant to read; and brilliant at unifying concepts siloed at computer science departments, neuroscience fields, cognitive sciences conferences and psychological publications in peer reviewed journals.

I swear, I read this book as slow as possible to savor every sentence. I wanted to tear through it fast - for its captivating nature in delightfully and deliberately explaining experiments on understanding the brain and its actions in the extensible world around us. I could pause in between paragraphs to awe at the above average quotidian polymathy the author displayed. It is truly one of those books you want to sit down with the mind that compiled it to have a decent chat on; and feel at liberty to ponder for more insights having tea or something.

At the end of day, this book left me more intelligible that I was before I picked it up. It not only taught me various researches on fathoming our minds, but also, it amplified my intuitions on areas of AI research I could embark on. All in all, it was truly nourishing. I am still in disbelief with how well it was written for a technical book, maybe the only one of its kind I have ever come across.

[Jacob Vorstrup Goldman · Rating 4 out of 5]

An enjoyable read with lots of interesting historical tidbits. Whilst the accusation that black-box probing of the brain under controlled experiments has quite limited explanatory power, and the book has a fantastic section on a paper that indicates just how little the data can be trusted with this method, I still think Summerfield succesfully manages to weave the insights and theories from neuroscience together with our current understanding of AI. And he also succesfully hammers home the fact just how unclear the idea of AGI and its desiderata are. The issue in the end is that whilst our AI and ML methods do show incredible promise, they are in fact also understood so poorly that our explanations of how they work is fraught with mindless theorizing of exactly the kind that is often leveled at neuroscientists. Is it world projections and compression via language a la Sutskever, or is it all an evolutionary procedure that is encapsulated through the loss function, such as for example Eran Malach provided evidence of in a paper on autoregressive next-token predictors just a week ago? Is the brain using gating and forget mechanisms or has a bidirectional transformer just replaced the brain? And that in the end is where the book falters a bit - detailing the interplay between two fluffy fields with an endless history of grandiose claims and hype ends up being too much Icaros and not enough Sisyphos. Humility is not something one associates with the old guard in either field, and in the end I suspect we will learn that it was all approximately wrong and exactly incorrect.

Faultridden musings:
- We have no idea lf language is necessary or sufficient for abstract reason, and the evidence provided for the former is completely useless
- Subconscious state-manipulation is completely not understood, Cormac McCarthy’s talk at Santa Fe on this very subject is very enjoyable
- the brain is a non-directed, endlessly loopy graph of tons of different components - its structure is endlessly more complex than anything we operate in the practice
- Vision is independent of actual brain workings, it is just a sensory aspect that happen to be processed in the brain and is important, but it is just as exciting to think about as touching of skin
- The Recife kids being good at street maths seem completely plausible, the genetic lottery book has more on that
- Intelligence might just be endless updating of some arbitrary function based on random mutations, it is so far the only procedure that has produced what we call intelligence
- Hebbian learning is too poorly understood

[Leon · Rating 5 out of 5]

Excellent book for someone with some background in AI or psychology and cognitive science hoping to better understand both fields, their deep historical connections, and current directions of research at their intersection. Summerfield is a brilliant researcher and truly an authoritative figure in this field. His writing is clear and insightful.

Highly recommend this book—it will not disappoint.

[Stefano Palminteri · Rating 5 out of 5]

It's clearly an academic book (ie it supposed some knowledge in cognitive science or neuroscience or artificial intelligence), but is a good one. From my point of view (cognitive scientists) I really found it useful for an overview of the intersection between my field and AI challanges.

[Jerome Han · Rating 5 out of 5]

An incredible overview of neuro/cogsci/AI research in the early 2020s.