Allen Newell and Herb Simon hired me to CMU about seven or so years before Geoff Hinton. I was hired into a Post Doc on a short telephone conversation with Allen when I said to him from UVa that I thought the 'serial basis for behavior' was wrong, and that it was all parallel, and I could prove it. I wound up writing an early perceptron-style Newellian production system called CAPS simulating the parallelism to predict eye movements in reading. I believe Terry's book is a must read for anyone coming into AI/ML/Big Data or Cognitive Neuroscience.
That said, I have a strong belief that Terry continues to underestimate what the science of human natural language, then called "psycholinguistics," has already taught us about computational cognitive neuroscience, and continues to teach us if we take the time to understand the direct observations on neural computation you can make that way. I would argue the mathematicians/statisticians now looking at over determined (high dimensionality) big data computations should look deeply into what we all can observe directly if properly educated in human natural language.
Similarly the NLP community (and much of Linguistics) is ~90%+ just delving into what we already know without such perspective either. But their success is their burden keeping major new theoretical work at bay. In any highly overdetermined system it is easy to find engineering ways to specific successes without much knowledge of what computational cognitive neuroscience and AI can learn if they only cared to do those direct observations that require little in the way of expensive or novel tools.
One of the cute stories of such overdetermination was my argument, now proven correct many times over, that getting cryptographically great random numbers of out mechanical spinning disk drives was a six week undergrad project, and every solution will be different. Enough motors, enough electronics, and electromagnetic storage, and voila, near perfect randomness can be found in thousands of places. Yet, even to this day, smart cryptographers warn it is not possible to get random numbers out of computing systems.
Understanding human cognitive computation out of the brain is easier if you use your brain to look in the right places in the right way.
Here is my review of Terry's book on the Deep Learning Revolution. More detail on this gap in what is otherwise an amazingly great and comprehensive story in his book:
Elon Musk should have an AI advanced NLP Conversation Group because his cars are going to need to reach Level 5 with customer satisfaction. Automation can only go so far without direct brain to "brain" communication. So should Google, Amazon, Apple, and others.
