Understanding 'A Thousand Brains', A Personal Review

Posted by : ()

Category : Reading

I recently finished reading Jeff Hawkins’s book, A Thousand Brains. I found this book to be incredibly enlightening, introducing a new framework for understanding the brain. This innovative perspective deepened my comprehension of the subject, transforming my previously vague notions about the brain into tangible, concrete concepts.

In this article, I will summarize the concepts presented in the first part of the book, starting from the level of the whole brain and gradually delving deeper until we reach the level of an individual neuron.

Our brain can be divided into two parts: the neocortex, which Hawkins refers to as “the new brain”, and everything else, which was referred to as “the old brain”.

The old brain, having evolved earlier in animals, is found in fish, birds, and reptiles. It is responsible for controlling emotions, desires, body movement, and the like.

On the other hand, the new brain evolved later and is exclusive to mammals. It enables the formation of ideas, languages, concepts, and more.

Structurally, the new brain envelops the old brain and is connected to it via numerous neural links, facilitating communication between the two.

If we were to flatten the neocortex, it would be, as Hawkins describes, “approximately the size of a large dinner napkin and twice as thick (about 2.5 mm)”. Imagine numerous circles on this napkin. Each circle corresponds to a structure called a cortical column, which measures about 1mm wide and 2.5mm high.

The main premise of the book is that this cortical column functions as the fundamental computational unit of the brain. The brain has approximately 150,000 cortical columns, each of them functioning like a mini-brain, capable of storing hundreds of different models and making predictions based on its input. This is why the book is titled “A Thousand Brains”.

A cortical column consists of multiple layers. Two important layers are discussed in the book. The upper layer receives sensory information while the lower layer receives movement information. Hawkins explains, “The basic flow of information goes as follows: A sensory input arrives and is represented by the neurons in the upper layer. This invokes the location in the lower layer that is associated with the input. When movement occurs, such as moving a finger, then the lower layer changes to the expected new location, which causes a prediction of the next input in the upper layer.”

This is analogous to the place cells and grid cells in the old brain, which provide animals with a sense of their location in the environment. However, whereas we have only one set of place cells and grid cells in the old brain, in the neocortex, we have a set of cortical place cells and cortical grid cells in every cortical column. When we interact with a coffee cup, for instance, the relevant cortical columns become active. Thanks to the cortical place cells and cortical grid cells, these columns can determine that the current sensory input corresponds to a certain point in the model, and when movement occurs, they can predict what sensory input will come next.

Cortical columns complement each other. Each object is stored “in many, but not too many, columns” simultaneously. This distributed storage system is both efficient and robust. Each perception or thought results from a consensus reached by the relevant columns through a process akin to voting. The ‘voting’ neurons act as representatives of a cortical column, broadcasting the column’s prediction through long-distance axons connected to other cortical columns.

It’s easy to see how this would work with everyday objects, like a coffee cup. However, this process also applies to objects we can’t see, such as a DNA molecule. We tend to visualize these invisible objects, allowing the cortical columns to form reference frames for these mental models.

Furthermore, the same method can handle abstract concepts. Instead of using a 2D or 3D physical space as a reference frame, a cortical column can create novel reference frames for abstract concepts. For example, when we think about political opinions, we might use the political spectrum to form a 1D space, with each opinion placed at a specific point on that spectrum. If we treat one opinion as an object, then the political spectrum represents one of the dimensions of the space in which the model lives. As our thoughts drift from one abstract concept to another, it’s as though we’re moving through one dimension of this abstract space, encountering new concepts along the way.

As Hawkins explains, “This is one reason that learning conceptual knowledge can be difficult. … Part of the learning is discovering what constitutes a good reference frame, including the number of dimensions. … Becoming an expert in a field of study requires discovering an effective framework to represent the associated data and facts.”

Cortical columns are composed of tens of thousands of neurons. According to the book, the model of the current artificial neural network doesn’t accurately reflect the structure and function of these neurons, and it proposes that we need a better model.

A neuron has many dendrites, with the vast majority of synapses—about 90%—located on these dendrites. When these synapses are activated, they trigger dendritic spikes that travel from the synapses to the cell body. Although these spikes don’t cause the neuron to fire, they make it easier and quicker for the neuron to fire. This process helps explain the phenomenon of priming in cognitive psychology. The remaining 10% of synapses, known as proximal synapses, are located near the cell body and are modeled as the input in the artificial neural network. In essence, the dendritic synapses prime the neuron, and the proximal synapses trigger the action potential. Without the dendritic synapses, the input would be too ambiguous for the neuron to generate a prediction.

The concepts outlined above comprise only the first part of the book, focusing on how the brain works. The latter part of the book, although equally fascinating and convincing, takes a more subjective approach, so I’ve chosen to exclude it from this summary.

To me, this framework is incredibly compelling. One of the most fascinating aspects of neuroscience is the sheer abundance of research papers containing a wealth of experimental data. Researchers measure various aspects of the brain and document their findings. This means that anyone can study neuroscience by reading these papers and forming their own theories about the brain. I’m excited to learn more about the brain and see how I can incorporate additional knowledge and information into this framework.

P.S. An interesting side note: The author of the book I read a week earlier, Max Tegmark, expresses significant concern about the existential risks introduced by artificial intelligence. Conversely, Jeff Hawkins sits on the other end of the spectrum, exuding optimism about the development of artificial intelligence. Both authors are very thoughtful and reasonable, yet it’s fascinating to see how they’ve arrived at vastly different opinions about the existential risk posed by AI.

Any feedback? We can discuss it under this Tweet.

About Sida Liu

I am currently a M.S. graduate student in Morphology, Evolution & Cognition Laboratory at University of Vermont. I am interested in artificial intelligence, artificial life, and artificial environment.

Follow @liusida
Useful Links