A Thousand Brains
December 3, 2021
Review by William P. Meyers

Site Search

Popular pages:

A Thousand Brains, A New Theory of Intelligence, by Jeff Hawkins

Basic Books, New York, 2021

Buy from Amazon (we get a cut): A Thousand Brains

How do humans think? How do we know the world through our senses, and survive in its complexity? How are we conscious? These questions have long haunted philosophers and scientists, with much answered in the last two centuries, but much remaining obscure. Jeff Hawkins has long worked on both neurology and artificial intelligence, has made some significant discoveries with his coworkers, and describes some of that in his latest book. I would call the book a primer, as it is an overview. The details of the discoveries are in papers cited in the book, which are on my To Do list, but this review is based on the book itself.

Artificial Intelligence v. Human Intelligence. Go back to 1820, when the Industrial Revolution was underway, precipitating the Slow Motion Apocalypse, and you would find an emphasis on speed. Machines that could make things quickly were appearing. A blacksmith or craftsperson who made something in an hour or day could be replaced by a machine that could spit out such an item in a few minutes. For instance spinning and weaving machines greatly expanded the production of cloth. No one then, or now, would claim that such a machine was more intelligent than a human being simply because it could work faster. But at the dawn of the modern era of Artificial Intelligence (AI), simply doing something faster was equated with intelligence, including intelligence superior to that of humans. There was a question as to what was mechanical [summing a large set of numbers for accounting or an engineering problem] and what seemed to be genuinely intelligent, like winning a chess game or being able to scan the text of an encyclopedia for an answer to a question. With the development of artificial neural networks in the 1980s, computers began to be able to quickly do things like image identification that were hard for humans to replicate. Was that a sign of intelligence, or just a mechanical trick? What distinguishes the two?

These computer feats have driven most AI researchers, neurologists and philosophers to look more closely at human intelligence and what it can accomplish. Any human baby can eventually learn a language, navigate the world, use tools, and make reasoned guesses about the future. A human can switch from chess to other games, and be adequate to very good at a number of them; a computer chess program is stuck in its ghetto. So perhaps intelligence should be viewed as a generalized ability to understand and act appropriately in the world, rather than being able to do some task (originally created by humans) faster than humans can. It is our brains, evolved over hundreds of millions of years, that allow for that.

On to brain science. Jeff Hawkins says that "We are left with a puzzle. The organ of intelligence, the neocortex, is divided into dozens of regions that do different things, but on the surface, they all look the same." The brain does have parts that more directly deal with sensory data as it comes in from the eyes, ears, skin, etc., and with outputs like arm and mouth muscle, but the cortex of the brain is where mammals gain their extraordinary flexibility of behavior. Take a real world chess game, between two humans. Players can see the board and recognize the pieces. They can move their muscles to shift a piece. They can think ahead, picturing the possible outcomes from moves. Yet at a moment's notice they may respond to audio inputs, or language, that may not be about the game. Does the brain have chess-piece recognition software, specialized chess-piece arm and finger movement routines, an interior gallery where future moves by the opponent are imagined? No, the higher functions are taking place in the thin, outer part of the brain using six layers of neurons that are pretty consistent over the whole surface.

Cortical Columns. Compare the quote above to this assertion: intelligence requires many things, but we know they are all done by neurons. Both statements are true, but not that helpful. What Hawkins asserts is the helpful division of analysis is the cortical column. Unlike the layers of the cortex, these columns do not jump out at you when observed with a microscope. However, the main directionality of neuron signals is up and down, though some impulses do spread horizontally. Various methods of brain imaging show that stimulus of particular sensory nerves results in specific columnar activity. The question then is how can a generalized cortical column be responsible, with the rest of the cortex and brain, for our rich understanding of the external world, and our rich mental life, like reasoning and imagination.

Reference Frames Hypothesis. Hawkins believes, based on some experimental evidence from neurology and computer modeling, that the cortical columns create maps, or reference frames, for information. While some areas of the cortex are more or less directly attached to neurons from various sense organs, or send signals to various muscles, the cortical columns begin as blank slates. As information, in the form of neural impulses, come in, they start creating maps, which can include features within the maps. This framework is so general it can handle visual information, discerning objects and actions, as well as building up the general scene, or space, that a person learns about. Yet it can also discern sounds, and then build up recognition of phonemes and words, and attach them to meanings of objects. It can learn where various parts of the body are in relationship to each other and to the overall scene. Once the learning is complete, with a bit of practice, picking up a baseball and pitching it into the strike zone is no big deal for most humans.

Orientation. Hawkins believes that each cortical column has three sets of cells: one that is the overall map, one that marks positions within the map, and one that tracks direction or orientation. What these mean is most clear when considering an animal learning to walk around a complex environment, but they generalize to all sorts of other forms of understanding. We understand that objects have orientations, which allows us to predict what we might see if we pick up a doll and turn it over, or try to hit a ball with a bat.

Images and Qualia. Both philosophers and AI researchers have simplified their tasks by thinking about objects (and actions) and qualities of objects (colors, textures). There is the tendency to think of representation in terms of pictures [like oil paintings] or images, or their equivalents for non-visual representations. I believe this leads to poor thinking in philosophy and computer science. Certainly, based on what we know of neurons and their synapses, there are not representations in the brain that are like the arrays of color dots that make up a color picture. My question, which I did not see Hawkins answer, is how his reference frames model explains our non-image view of the world, which I prefer to call a scene view [as did Nobel Prize winner Gerald Edelman]. The scene before me now includes a computer screen that resembles an analog picture, but it is in a three dimensional scene context where my eyes and hands are in front of the screen and, looking out my window just past my screen, I can see trees, the sky, and if I am lucky birds or squirrels. When I hear a sound, it comes from somewhere in the scene. This representation problem is related to the issue of qualia, which Hawkins does devote a couple of pages to. As far as we know everything in our brain is just electro-chemical spikes. Why are we conscious of some spikes as red, some as blue, some as bird songs, some as language, etc.? Hawkins take a stab at it, as many philosophers and others have. If you want a big research topic to work on, qualia sits there, mysterious and largely undefeated. Some philosophers have challenged the very concept of qualia, but the issue of neuron spikes v. what we are conscious of remains even it you dance around qualia.

Consciousness I. The book does not dwell very much on issues of consciousness, but I will because that is what interests me most. It helps to divide consciousness into two types. We call a human, mammal, or bird (and sometimes other animals) conscious when they act as if they are aware of the exterior world. A mouse sees a cat and runs to hide. That we call Consciousness I. It can be explained by tracing the firing of sensory nerves along their various paths and connections to the parts of the mouse brain that interpret the senses and make decisions, including to order the muscles to run for cover.

Consciousness II is more of a philosophical problem. Does the mouse see the cat in roughly the sense that I see the cat (does the mouse have a qualia issue?). Is the mouse aware that it sees the cat? Awareness is Consciousness II. Some philosophers believe language is needed for Consciousness II, but I am not sure of that. There is a bad tendency of philosophers, even Ludwig Wittgenstein, to be overly impressed by their own language and reasoning skills. Every attempt I have seen by philosophers to illuminate the mind-body problem falls apart when the various aspects of Consciousness II are considered. Hawkins reasonably asks if machines can become conscious, rather than exactly how human Consciousness II works. "I expect that a similar change in attitude will occur with consciousness. At some point in the future, we will accept that any system that learns a model of the world, continuously remembers the states of that model, and recalls remembered states will be conscious. There will be remaining unanswered questions, but consciousness will no longer be talked about as the hard problem." [emphasis mine] I, on the other hand, suspect it will prove to be the hardest problem of them all.

A Good Basis for Further Development. I am impressed by Jeff Hawkins and his team's work on cortical columns and some of the harder AI and human brain issues like how reference frames are created and populated. I would like to learn more about what drives the decision making in these reference frames, because having a map and deciding to take a look around are not automatically an intelligent behavior. The simple yet complex chessboard reference frame populated by chess pieces and the much more complicated world of even the most primitive hunter-gatherer are both challenges for humans. The more we know about the actual mechanisms involved, the better we should be at meeting future challenges.

Mind Body Problem Left UnsolvedIt should surprise no one familiar with the fields that A Thousand Brains leaves the mind-body problem unsolved. Hawkins is aware of it, but it is not particularly on his agenda. As Patricia Churchland and others have pointed out, we may have a wealth of detail to learn about the brain before we find something that triggers the right insights to understand the mind-body problem and provide a comprehensible solution. I am not convinced the cortex in itself is the right or sole place to look for insights into the problem. Other, older areas of the brain may also have developed in humans that give us capabilities beyond apes and other mammals. The whole brain needs to be examined, but the cortical columns, if they are indeed working mainly by generating reference frames, certainly are important to the overall project.

I highly recommend A Thousand Brains to anyone interested in AI, the brain, neurology, philosophy, psychology or human civilization.