A thousand brains by Jeff Hawkings. This book is quite special in many regards. The first thing I have to say about it is that my first impression was not good, but then when I went past a bit over half the book then it fell on to me that this book is visionary.
I went almost half way through the lecture of this book and then I almost gave up, not because I did not find it interesting but because I felt it was too long and I thought a short essay or just an article would be enough. I am glad I didn’t stop reading.
This book proposes a new theory of intelligence and explains how by understanding how intelligence works eventually we will be able to create true intelligent machines.
The basic hypothesis of the book is a new approach to the understanding of the brain. We used to believe that our brain is divided into different functional areas of specialization corresponding to the different sensorial stimuli: vision, listening, abstract thinking etc. And this belief is partially true. For example the part of the brain located in the occipital lobe at the back of our head is devoted to process vision. We know this because people with head trauma on the back of their head often lose their vision. And so it is true that the brain has specialized regions, however the whole neocortex contains more or less the same types of neurons. In other words the brain has areas of specialized functions but these areas have similar constituents.
One of the main ideas the author wants to put forward in this book is the concept of the cortical column, which occupies about one square millimeter of the neocortex and each column is divided into several hundred “minicolumns”. Each minicolumn contains about one hundred neurons. All cortical columns operate through a common principle and what a cortical column does is process sensory input. The processing of information is basically the same in any cortical column, regardless of their location in the neocortex. Then the same portion of brain tissue is able to process vision or language or any other kind of brain process. In this regard the neocortex is adaptive.
Our brain is a predictive machine that anticipates our response to diverse stimuli of the world. And the purpose of each cortical column is to process the external stimuli to predict how to react to various kinds of stimuli. One example could be predicting the next footstep as we walk or how to handle a cup of hot tea to avoid dropping it or getting burned. The cortical columns also handle abstract thinking.
In order to make predictions the brain needs to build a model of the world. Let us consider a car as a simple example. If we imagine a car, our brain has a model that has learned from experience and from that model we can imagine the car from different angles or in different positions.
So there it is: the most relevant aspect of the brain to understand the concept of intelligence is the neocortex and how it operates. The neocortex is made out of cortical columns and the goal of these cortical columns is to process sensory data to produce a model of the world.
The second part of the book is visionary. Based on the understanding of intelligence and how the brain works, the author proposes that true intelligent machines could exist in the near future.
Artificial intelligence (AI) as we know in the present time is often exemplified by neural networks which although have impressive power of learning, this learning is too specialized to be considered intelligence. For example, a machine executing a computer code “A” might be able to perform face recognition and another code “B” could be capable of making weather predictions. However computer code “B” would not be capable of doing what code “A” does. True general intelligence is adaptable and multipurpose. Machines possessing true artificial intelligence should be able to learn how to do anything by themselves.
It is necessary to rethink the foundations of AI in order to design true intelligent machines that some day will be able to use external sensory input to construct habitats in unknown environments on other planets or to understand molecular forces to reconfigure protein structures by using chemical sensory inputs, among many other possible future applications.
My last comment on this book is that although the idea of artificial cortical columns as a framework for an all-purpose general artificial intelligence is really appealing, it still needs to be fully verified as a viable theory.