A Thousand Brains: A New Theory of Intelligence

by Jeff Hawkins

We are creating powerful technologies that can fundamentally alter our planet, manipulate biology, and soon, create machines that are smarter than we are. But we still possess the primitive behaviors that got us to this point. This combination is the true existential risk that we must address. If we are willing to embrace intelligence and knowledge as what defines us, instead of our genes, then perhaps we can create a future that is longer lasting and has a more noble purpose.

I find it amazing that the only thing in the universe that knows the universe exists is the three-pound mass of cells floating in our heads.

When the neocortex wants to do something, it sends a signal to the old brain, in a sense asking the old brain to do its bidding.

Neurons in some layers make long-distance horizontal connections, but most of the connections are vertical. This means that information arriving in a region of the neocortex moves mostly up and down between the layers before being sent elsewhere.

We now know this description is misleading. In every region they have examined, scientists have found cells that project to some part of the old brain related to movement. For example, the visual regions that get input from the eyes send a signal down to the part of the old brain responsible for moving the eyes.

There are no pure motor regions and no pure sensory regions.

Mountcastle proposed that the reason the regions look similar is that they are all doing the same thing. What makes them different is not their intrinsic function but what they are connected to.

Cortical columns are not visible under a microscope. With a few exceptions, there are no visible boundaries between them. Scientists know they exist because all the cells in one column will respond to the same part of the retina, or the same patch of skin, but then cells in the next column will all respond to a different part of the retina or a different patch of skin. This grouping of responses is what defines a column.

Finally, there is the argument of extreme flexibility. Humans can do many things for which there was no evolutionary pressure. For example, our brains did not evolve to program computers or make ice cream—both are recent inventions. The fact that we can do these things tells us that the brain relies on a general-purpose method of learning. To me, this last argument is the most compelling. Being able to learn practically anything requires the brain to work on a universal principle.

Prediction was a ubiquitous function of the neocortex.

When the brain’s predictions are verified, that means the brain’s model of the world is accurate. A mis-prediction causes you to attend to the error and update the model.

We can think of the neocortex as starting life having some built-in assumptions about the world but knowing nothing in particular.

Heavy emph. On nurture

The big insight I had was that dendrite spikes are predictions. A dendrite spike occurs when a set of synapses close to each other on a distal dendrite get input at the same time, and it means that the neuron has recognized a pattern of activity in some other neurons. When the pattern of activity is detected, it creates a dendrite spike, which raises the voltage at the cell body, putting the cell into what we call a predictive state. The neuron is then primed to spike. It is similar to how a runner who hears “Ready, set…” is primed to start running. If a neuron in a predictive state subsequently gets enough proximal input to create an action potential spike, then the cell spikes a little bit sooner than it would have if the neuron was not in a predictive state.

If their preferred input arrives, they all want to start spiking. However, if one or more of the neurons are in the predictive state, our theory says, only those neurons spike and the other neurons are inhibited. Thus, when an input arrives that is unexpected, multiple neurons fire at once. If the input is predicted, then only the predictive-state neurons become active. This is a common observation about the neocortex: unexpected inputs cause a lot more activity than expected ones.

At first, it was hard for us to imagine how neurons could represent something like x, y, and z coordinates. But even more puzzling was that neurons could attach a reference frame to an object like a coffee cup. The cup’s reference frame is relative to the cup; therefore, the reference frame must move with the cup.

Since the complex circuitry in every cortical column is similar, locations and reference frames must be universal properties of the neocortex.

As I will explain in coming chapters, the brain builds models of the world by associating sensory input with locations in reference frames.

The likelihood that a solution is correct increases exponentially with the number of constraints it satisfies. It is like solving a crossword puzzle:

With these additions, we showed that a single cortical column could learn the three-dimensional shape of objects by sensing and moving and sensing and moving.

They discovered what are now called place cells: neurons that fire every time the rat is in a particular location in a particular environment.

They discovered what are now called grid cells, which fire at multiple locations in an environment.

Every time a rat enters an environment, the grid cells establish a reference frame. If it is a novel environment, the grid cells create a new reference frame. If the rat recognizes the environment, the grid cells reestablish the previously used reference frame.

Again, to learn a complete model of something you need both grid cells and place cells. Grid cells create a reference frame to specify locations and plan movements. But you also need sensed information, represented by place cells, to associate sensory input with locations in the reference frame.

Grid cells and place cells in the old brain mostly track the location of one thing: the body. They know where the body is in its current environment. The neocortex, on the other hand, has about 150,000 copies of this circuit, one per cortical column. Therefore, the neocortex tracks thousands of locations simultaneously. For example, each small patch of your skin and each small patch of your retina has its own reference frame in the neocortex. Your five fingertips touching a cup are like five rats exploring a box.

In the paper-map analogy, I described looking at the map squares one at a time. This could take a lot of time if you had many maps. Neurons, however, use what is called associative memory. The details are not important here, but it allows neurons to search though all the map squares at once. Neurons take the same amount of time to search through a thousand maps as to search through one.

The upper layer is roughly equivalent to place cells and the lower layer is roughly equivalent to grid cells.

Put another way, an object such as a coffee cup is defined by a set of observed features (upper layer) associated with a set of locations on the cup (lower layer). If you know the feature, then you can determine the location. If you know the location, you can predict the feature.

Vernon Mountcastle proposed that every column in the neocortex performs the same basic function. For this to be true, then, language and other high-level cognitive abilities are, at some fundamental level, the same as seeing, touching, and hearing. This is not obvious. Reading Shakespeare does not seem similar to picking up a coffee cup, but that is the implication of Mountcastle’s proposal.

Of course, nobody has ever directly seen or touched a DNA molecule. We can’t because they are too small. To organize our knowledge of DNA molecules, we make pictures as if we could see them and models as if we could touch them. This allows us to store our knowledge of DNA molecules in reference frames—just like our knowledge of coffee cups.

The hypothesis I explore in this chapter is that the brain arranges all knowledge using reference frames, and that thinking is a form of moving. Thinking occurs when we activate successive locations in reference frames.

But it is not necessary to abandon Mountcastle’s premise. We have proposed a simple explanation for why some columns are what columns and some are where columns. Cortical grid cells in what columns attach reference frames to objects. Cortical grid cells in where columns attach reference frames to your body.

It is likely that columns in the neocortex don’t have a preconceived notion of what kind of reference frame they should use. When a column learns a model of something, part of the learning is discovering what is a good reference frame, including the number of dimensions.

The research team further showed that when the subjects thought about birds, they were mentally “moving” through the map of birds in the same way you can mentally move through the map of your house.

Our thoughts are continually changing, but they are not random. What we think next depends on which direction we mentally move through a reference frame,

My point is that becoming an expert in a field of study requires discovering a good framework to represent the associated data and facts. There may not be a correct reference frame, and two individuals might arrange the facts differently. Discovering a useful reference frame is the most difficult part of learning, even though most of the time we are not consciously aware of it.

Within a cortical column, the previously learned coffee cup is defined by a reference frame. The previously learned logo is also defined by a reference frame. To learn the coffee cup with the logo, the column creates a new reference frame, in which it stores two things: a link to the reference frame of the previously learned cup and a link to the reference frame of the previously learned logo. The brain can do this rapidly, with just a few additional synapses. This is a bit like using hyperlinks in a text document.

Reference frames in the old brain learn maps of environments. Reference frames in the what columns of the neocortex learn maps of physical objects. Reference frames in the where columns of the neocortex learn maps of the space around our body. And, finally, reference frames in the non-sensory columns of the neocortex learn maps of concepts.

This is why we call it the Thousand Brains Theory: knowledge of any particular item is distributed among thousands of complementary models.

The basic idea of how columns can vote is not complicated. Using its long-range connections, a column broadcasts what it thinks it is observing. Often a column will be uncertain, in which case its neurons will send multiple possibilities at the same time. Simultaneously, the column receives projections from other columns representing their guesses. The most common guesses suppress the least common ones until the entire network settles on one answer.

Instead of the neocortex using hierarchy to assemble features into a recognized object, the neocortex uses hierarchy to assemble objects into more complex objects.

The difficult part of knowledge is not stating a fact, but representing that fact in a useful way.

When a neuron learns a new pattern, it forms new synapses on one dendrite branch. The new synapses don’t affect previously learned ones on other branches. Thus, learning something new doesn’t force the neuron to forget or modify something it learned earlier. The artificial neurons used in today’s AI systems don’t have this ability. This is one reason they can’t learn continuously.

How does the brain do it? The unit of processing in the neocortex is the cortical column. Each column is a complete sensory-motor system—that is, it gets inputs and it can generate behaviors. With every movement, a column predicts what its next input will be. Prediction is how a column tests and updates its model.

We are intelligent not because we can do one thing particularly well, but because we can learn to do practically anything.

The active neurons in the brain at some moments represent our present experience, and at other moments represent a previous experience or a previous thought. It is this accessibility of the past—the ability to jump back in time and slide forward again to the present—that gives us our sense of presence and awareness. If we couldn’t replay our recent thoughts and experiences, then we would be unaware that we are alive.

The redness of the fire truck is a fabrication of the brain—it is a property of the brain’s model of surfaces, not a property of light per se.

Fear of death and sorrow for loss are not required ingredients for a machine to be conscious or intelligent. Unless we go out of our way to give machines equivalent fears and emotions, they will not care at all if they are shut down, disassembled, or scrapped.

We can’t know the future, and therefore we can’t anticipate all the risks associated with machine intelligence, just as we can’t anticipate all the risks for any other new technology. But as we go forward and debate the risks versus the rewards of machine intelligence, I recommend acknowledging the distinction between three things: replication, motivations, and intelligence.

Replication: Anything that is capable of self-replication is dangerous. Humanity could be wiped out by a biological virus. A computer virus could bring down the internet. Intelligent machines will not have the ability or the desire to self-replicate unless humans go to great lengths to make it so.

Motivations: Biological motivations and drives are a consequence of evolution. Evolution discovered that animals with certain drives replicated better than other animals. A machine that is not replicating or evolving will not s...