On 'On Intelligence'
To distinguish between a cat and a dog, a computer needs to execute billions of steps. A human can tell the difference in less than 100 steps. To give you an idea, 100 steps are not enough to move even a character on a computer's display.
How is this possible?
This difference in abilities can be attributed to the way humans process information vs how computers do it. Humans don't actually compute solutions like computers do. Instead, they recall solutions from memory. The answers are stored in the memory a long time ago from experience. It does not take much time for the solutions to be retrieved. Neurons in the cortex constitute memory themselves. In fact, the whole cortex is a memory system. Humans use a combination of memory recall and incorporation of the specifics of the moment- a process called prediction -to compute solutions.
Pointing out a fallacy in the Turing test, the author argues that prediction, and not behavior, characterizes intelligence. He believes the neocortex is the part of the brain responsible for intelligence in humans.
How is memory stored so that it can be recalled appropriately and applied to new situations?
Memory is stored in invariant representations. Humans store memories in the form of invariant patterns. In any situation, the brain identifies patterns in terms of relativeness. For example, to remember the face of a friend, a particular image of the friend at a particular instant is not stored. "What makes the face recognizable are its relative dimensions, relative colors, and relative proportions. There are "spatial intervals" between the features of her face just as there are "pitch intervals" between the notes of a song. Her face is wide relative to her eyes. Her nose is short relative to the width of her eyes. The color of her hair and the color of her eyes have a similar relative relationship that stays constant even though in different lighting conditions their absolute colors change significantly. When you memorized her face, you memorized these relative attributes."
Memory is recalled in an associative manner. Whenever we experience something, memories similar to what we are experiencing(association) are recalled. Based on these recollections, predictions are made on the sequence of events that will follow. When the sequences go against expectations, either old memories are updated or new memories are created.
Other Key Ideas Proposed
The author argues that humans are much more than just intelligent machines. They are biological beings with reproductive urges, hunger, emotions. To build an intelligent machine but not a human like one, you just need to focus on the part of the brain responsible for intelligence in humans - the neocortex. Consequently, the model does not need to be embodied.
The neocortex is constituted of 6 layers. Input to the layers varies in decreasing order of variability. For example, while looking at a pen, the lower layers receive very noisy information based on saccading on the eyes. As information passes into the higher layers, it is in the form of edges, shapes and complete images.
Scientists have identified through experiments areas of the neocortex associated with specific functionalities such as processing audio, visual data. However, it has been found that the neocortex is plastic in nature. These functional areas, when connected to other sensory regions, learn to process the input from those sensory regions.
We do not have an absolute perception of the world. The brains knows the world through a set of senses which can perceive parts of the absolute world.
All modalities of input such as auditory, visual etc. reach the neocortex in the same format of representation.
The similarity in the structure of the neocortex gives rise to the notion that neurons at all locations maybe performing the same basic operation - a single cortical algorithm
More than the size of the neocortex, the dominance of the neocortex on other parts of the brain is a characteristic of intelligence. The primary function of the neocortex is to make continuous predictions about the world around us.
In creating an intelligent machine, if the machine is not programmed to have intrinsic motivation such as hunger, curiosity etc, on what basis will it explore the world to to build memories essential for prediction? If the programmer decides the sensory input for the agent, is the agent really intelligent? Not everything we experience is stored into memory. What gets stored depends on the attention we pay. How will the model confer attention on certain aspects of its environment without intrinsic motivation?
Given my reservations above, the author's ideas about memory which is stored in patterns, is invariant and auto-associative is are useful for modeling memory. The idea of a single cortical algorithm is seems plausible, although discovering it will be a long and arduous journey.
A very new idea I read for the first time is that all sensory input eventually reaches the neocortex in the same representation format. Maybe this is something we can think of while designing sensory inputs to models.