Book review of Marvin Minsky

Marvin Minsky:

"The Emotion Machine" (Simon & Schuster, 2006)
(Copyright © 2015 Piero Scaruffi | Terms of use )

This book summarizes all of the cognitive ideas of Minsky, from frames to K-lines, and some new one, and looks like an update thirty years later of Minsky's "The Society of Mind" (1985) . It is, however, poorly organized, and the arguments sound mostly weak and outdated (the book came out the same year that Hinton discovered "deep learning"). Minsky's ideas are pure guessing of psychological and neurological facts with little or no awareness of contemporary literature. He is clearly still influenced by the old "symbolic" or "knowledge-based" school of A.I. and even takes a couple of swings at neural networks ("you cannot learn things that you can't represent", "i see their popularity as having retarded the search for higher-level ideas"). Minsky mostly relishes filling his chapters with endless quotations from philosophers, scientists, psychologists, artists, etc.
The book opens with a discussion of goals. He argues that our higher-end goals are inherited from the people we are attached to, a form of imprinting. We then abstract those goals into our ideals and moral codes. He then organizes our mental resources into six levels starting with instinctive reactions and peaking with self-conscious emotions.
The best section of the book is when Minsky takes issue with those who simplify the brain as being a network. If it were just one network, it would not have survived natural selection: the bigger the size of a system the lower its performance (he doesn't prove this statement). He argues that, for a complex system to survive, it must maintain a balance of division and unity: its parts cannot become too entangled but cannot be too separated either. The only evidence for this theorem that he presents is the fact that all bodies are made of many organs, not just one catch-all organ. I find this idea more plausible than the idea that the brain is one giant neural network.
On the other hand, when he discusses consciousness, his arguments sound so naive and confusing that i wonder whether Minsky was a conscious being... I do agree that our mental life is not one monolithic phenomenon but instead a collection of different mental processes. How this relates to the problem of consciousness is not clear. I suspect he missed at least two decades of neuroscience studies (although he mentions Damasio a couple of times).
He also reprises old concepts of "Society of Minds" such as the "simulus", a combination of stimulus and simulate that, in his opinion, explains how the brain constructs an image. He also draws a simplistic "prediction machine" (so simplistic that i had to roll my eyes).
The chapter on common sense is the longest and, potentially, the most valuable, but it's mostly old theories of knowledge-based systems (that may or may have not yielded useful clues to build intelligent systems): semantic networks, scripts, frames, Winston's different networks, knowledge-lines, etc. It's a hodgepodge of old techniques that is not particularly enlightening.
His "critic-selector" model of mind is a generalization of the old production systems, the systems based on rules of the "if then" kind. The critic recognizes the type of the situation and the selector activates one of the possible ways of thinking. He speculates that before the evolution of modern humans there was no selector: a critic generated only one kind of thinking, the genetically-programmed one. Now we have the luxury of selecting one of the many ways that we can think about a problem, and we can even invent new selectors and new critics. We have many different ways to achieve a goal. This multiple `ways to think' capability requires that we think of an object or a situation in many different ways: a car is a vehicle, is an investment and is a cause of death. We think in terms of "parallel analogy" or "panalogy".
He comes up with a hierarchy of representations that starts at the bottom with "micronemes" (a variant of the "microfeatures" proposed by David Waltz and Jordan Pollack in 1985), then neural networks, then K-lines, then semantic networks, then frames, and at the top narrative stories. Our brain must be born with primitive forms of semantic networks, frames and K-lines. Brains cannot invent new types of representation because a representation is useless without the skills to use it.
The last chapter on the "self" is as confusing and amateurish as the chapter on "consciousness". Worse: here he tries to link his theories of the six levels with Freud's theory.
The most powerful ideas in this book are: 1. we have multiple ways to represent knowledge; 2. we have multiple ways to think about our knowledge; 3. the brain is made of many different organs, and each one probably evolved to compensate for the deficiencies of the others.