Precis: On Intelligence
On Intelligence By Jeff Hawkins with Sandra Blakeslee 2004
Hawkins’s claim to fame is that he made Graffiti, the quirky handwriting-recognition system for PalmPilot. In Graffiti, users write highly stylized versions of each character, which are extra easy for the computer to recognize. He made the people simplify themselves so the computer could take their direction better.
This book is not about his work, but his hobby: finding a neurological basis of intelligence to use as a model for artificial intelligence. Cool idea, right?
His central claim: you can’t simulate something you don’t understand. I had to wrestle with this for a bit to accept it. (But now don’t anymore.)
Can you simulate something you don’t understand? Can you just copy the position and basic function of every neuron, and have a brain? Can an actor playing a doctor perform surgery?
Hawkins would say no. You have to understand how the thing really works to simulate them right. You can make a crappy copy of a thing, like SimCity where it’s “a city.” But it’s not really a city. And that’s how AI is when we don’t understand what intelligence really is. Even if you copied every neuron, you have to understand what they do to understand how they work.
So, intelligence. It arises from human brains. If you don’t understand the brain, you can’t simulate intelligence. Thus, for Hawkins, all existing AI is not intelligent. The necessary first step to AI is an all-encompassing theory of how intelligence is produced by the brain.
As many reviewers have noted, Hawkins prefers to present as an authority, rather than as a student, and his gifted co-author seems to have had little creative control in the project. There are, in fact, many other approaches to AI that ignore human brains completely. That is the basis of the Turing test, Eliza (the primitive chatbot), machine learning and neural networks. These models all work in practice, but not, for Hawkins, in theory.
Hawkins wants to define intelligence and starts with the fun claim that intelligence is what the brain does. So, if you can understand what the brain is doing, you will have intelligence ipso facto. In a rich white male engineer move I see often in California, Hawkins sees existing human cultural constructions as meaningless and in need of more technical clarity. “Intelligence” needs to be specc’d! I’d point out that get a lot done with our funny animal chatter words, especially “intelligent,” such as sell copies of this book, help us recommend things to each other, gather up the right friends for board game night, and give this successful businessman something interesting to chase after. But it is not enough. He needs a specification. Thus the chase. And what a chase it is!
The neocortex is the only part of the brain that matters for understanding intelligence. You can still think without some of the other parts, even if you need them for emotions or friendship or whatever. Behavior is not a critical part of intelligence, as you can be intelligent with your eyes closed in the dark without moving!
The neocortex is a sheet six layers deep of neurons. Signals go down the neurons, coming in from the senses, but also go back up, adding context and interpretation to sensation. This process of interpretation is basically one of comparing a vast store of tiny memories to the diachronic pattern of a sensation. He has a fantastic, tiny, example of this: how do you recognize gravel by touch? You have to move around on it, to sense the pattern of gravel pieces. Melodies and rhythms are naturally like this too. There is an element of time to perception, and memories are tested and compared over time.
Hawkins delivers lots of fun, digestible pop brain science (maybe thanks to Blakeslee). One major fulcrum of his inferential lift is that humans can do lots of things in under half a second, but during that time a signal can only travel through 100 neurons. How the hell does the brain do that? From his background in computing, he is easily able to argue that typical software simulations of intelligence could never match this. To catch a ball (his example), a computer could pick apart a series of images, infer trajectory and size (maybe weight?) and calculate how to move a hand to grab it. (Video of a robot doing this.) But this takes more than 100 component steps. So the brain cannot use this technique. Instead, Hawkins surmises, the brain must compare a large set of similar memories, trying to fit what to do against what it is seeing, basically copying the other times it tried to catch a ball. If the ball is larger or darker or faster than before, the brain must be able to compensate for this very quickly. Interesting concept about memory’s role in active processing!
In general, our senses are crap, he points out. It is the neocortex that gives us the illusion we experience the world as it really is. Our eyes are like one megapixel cameras with a thick fish eye lens, pointing and focusing on new things real fast. Only the fovea (at the very center) has much detail, and it moves around in unconscious saccades, targeting what we consider the important bits of things, such as eyes and faces. Vision, he claims, is better understood as three senses: luminance, color, and motion. Hearing is also very deceptive, telling us we hear the same melody even when, in different keys, there are actually no distinct sounds in common between different renditions of a tune.
Hawkins argues that the brain has the same architecture for each sensory region and, in this experiment one time, you can actually hook up a sensory organ to a different part of the brain and the mind learns to interpret the signal just fine! Thus all senses are alike, to the neocortex, and the interpretation of sensation is fundamental to intelligence.
The concept of intelligence explained here reminds me of Hercule Poirot, the famous fictional detective, whose special power is simply to notice anything that is out of the ordinary. Hawkins argues that this is the usual way of the mind, memorizing patterns of things and making a fuss when anything is distinctive.
Hawkins goes on about the columnar architecture for many pages, reframing prediction as imagination and discussing the lateral connections between “functional units.” I lost interest because I’m not going to build an AI on his model. I just wanted to learn about an interesting concept of intelligence, and I did.
Remember that the thesis of this book is that this model of intelligence gives us the basis for real artificial intelligence — software that emulates intelligence, rather than simulates it (my words, not his).
In the final sections of the book, Hawkins returns to AI. He is very skeptical that AI will ever be “sentient” or self-replicating or any of the other things we often imagine. It certainly won’t be human, as it will lack emotional systems and the bodily experience of being human, not to mention socialization. It also won’t be really useful as a butler or C3PO type friend, mostly for economic reasons, he suggests.
Instead, he believes AI will have some advantages over humans minds and disadvantages. He focuses on the advantages. AI will have more capacity, able to do more of whatever it can do than a human could for longer at larger scale. Similarly, it will be faster — this is already true of almost everything computers can “think” through. AI will also be more replicable than humans, able to be manufactured at scale with pre-established memory (thus also programming?) or without! Finally, AI will have better sensory perception than humans, able to integrate better sensory inputs with larger neocortex-like processing.
These statements are very reasonable and I like Hawkins on AI more than I like him on neuroscience, it turns out. Human sensation still sets the standard for what we believe the world really is, but this should become less true in the future. In a way, this upsets the basis of humanism, as the old “man is the measure of all things” turns into “human taste is most worth serving, but things are measured with machines.” Human capacity for cognition has long set the limit on many things that were just too damn hard to think through any better, but that is already changing in places where there is money to be made by having computers in place of smarties. Replicability also means a risk of drowning out other things, of an overwhelming surfeit of machine thought and output over human. And speed is already the basis of the drone revolution, with auto-stabilization making quad-copter and other designs that didn’t work for human pilots into a solid platform for robots to be flying about now as they please. Some very specialized robots can even catch a ball.
But his theory of mind is productive, giving us a good chase around topics of memory and perception, and providing some ideas for a different approach to AI. I don’t have much confidence in AI built on these principles of how the brain works, but then most AI models are pretty weak.
To me, the great irony of this work is that Graffiti, his opus, showed that you don’t really have to understand how people work to make computers useful to them. But here, as if in some kind of act of vengeance, Hawkins insists that you must understand how humans think in order to make machines that think too.
I came to this book looking to learn about intelligence as a relative ability between different people, and instead got this rather unrelated wild ride which I mostly devoured in an airplane ride wasted on jet lag and coffee. I think Hawkins does decompose intelligence to a number of related technologies, and it may be possible to understand these technologies as more or less in evidence for different people, or, hell, for machines too.
