On Intelligence is a book written by Jeff Hawkins that proposes a new approach to understanding intelligence and building intelligent systems. Jeff Hawkins and his team of researchers have over the years accumulated several research papers in the field of Neuroscience and uncovered a hypothesis about the working of the human brain. He touches on many subjects, such as the brains ability to recognize patterns and build invariant representations of the world from them, but the main idea of the book is that prediction is the key thing that makes the brain intelligent.
Intelligence and learning lead to better prediction, a mistake in the prediction leads to learning. He argues that a single cell organism is also intelligent, because it can make predictions about its environment by using its senses and its “programming” embedded in its DNA. Going up the hierarchy, reptiles are more intelligent because they can accomplish more complicated tasks. Mammals though are even more intelligent due to the fact that they can not only make predictions about the world, but because these predictions can also change with experience. They are not static (at least not all of them), a result from evolutionary “programming”, but they are ever changing during the animal’s lifetime. The point here is that while prediction is the key to intelligence, it requires a complicated memory system in order for it to work. The author calls it the “memory-prediction framework”. Humans are at the top of the “intelligence pyramid” because of their ability to pass on information from their memories and create such memories in others without the need for the other to have actually experienced the memory directly.
The following figure shows how the information of sight is processed in the human brain. The eyes are connected to the temporal lobe of the brain through optic nerves. The information gets processed here along with the spatial information obtained from the entrohinal hippocapal and from the hypothalamus region. We can clearly see a hierarchical architecture. At each of the sub-region V1,V2,V4 and IT different levels of abstraction of the input happens to understand the observed spatial-temporal pattern. This helps store sequence of recurring patterns and recall them auto-associatively by having them in an invariant form and in a hierarchy.
Baddeley’s Model of Working Memory is a model of human memory proposed by Alan Baddeley and Graham Hitch in 1974, in an attempt to present a more accurate model of primary memory. Working memory splits primary memory into multiple components, rather than considering it to be a single, unified construct. This is the main model that first introduced the idea of hierarchy in the memory formation and recall.
HTM is a hypothesis that explains the working of the neocortex, the largest part of the brain and how it can be implemented to build a truly intelligent system. The book dwells upon the idea of intelligence of the human brain and the true working of the neurons. The key is to understand the working of neurons in the outermost layer of the brain, the Neocortex. There are about six columned layers in the human neocortex, highest in any mammal. The neurons are arranged in a set of column vertically and are connected horizontally across a layer. This can mostly be accredited to Vernon Benjamin Mountcastle was Professor Emeritus of Neuroscience at Johns Hopkins University whose work revealed the true nature and working principles of a neuron in the paper Brain mechanisms for directed attention.
Another paper that really help build and prove the hypothesis of HTM is Memory, navigation and theta rhythm in the hippocampal-entorhinal system by György Buzsáki & Edvard I Moser, which talks about the discovery of grid cells, a special type of neurons present at every layer that helps locate the current world position of once self, think about navigating to a destination and form a basic model of the world. It is proposed here that memory developed due to the need of organism’s movement through its surroundings. The paper speaks about how there are two types of views that help form the whole world model. Allocentric view is perspective with respect to a particular landmark in the surrounding. Egocentric view is perspective from one-self. By looking at these discoveries Jeff came to conclusion that there has to be a simple algorithm that is being implemented at the fundamental level in the neocortex to understand and form complex association of different sensory inputs into a virtual model and recognize already observed patterns and their combinations.
Each system provides a unique reference point from the grid cells, a combination of these reference cells gives a fairly accurate representation of current location in the world.
By truly understanding the working of this Cortical Learning algorithm we can implement it to have a brain like learning, remembering and predicting capabilities on a system.
Further, he discusses many more interesting examples that give an accurate insight into the brain’s working and helps define Intelligence. Now that we have an overview of the book we can move on to actually breaking down the Algorithm