How a New Understanding of the Brain Will Lead to the Creation of Truly Intelligent Machines
By: Jeff Hawkins
How does the brain work, specifically the (neo)cortex, and what is intelligence.
Its central theorem is the memory-prediction framework, which explains how the brain (and potentially other complex systems) are able to display intelligent behavior.
Memory systems are fed ongoing streams of sensory inputs, which are stored in certain types of patterns. Intelligence then is the capacity to receive these inputs from the senses, compare and combine them with stored patterns to predict what happens next.
As long as the world around us has sufficient structure and interesting patterns to observe, brains (and other systems) will have an opportunity to intelligently exploit their environments by detecting and storing patterns and trying to predict what comes next.
Short and well written. Develops interesting theories based on clear and simple definitions.
The section explaining how neuronal layers in the cortex interact in storing and retrieving patterns is particularly interesting (the signalling back and forth, the hierarchical layers, the flows of signals up and down the hierarchy as patterns are stored, learned and retrieved).
- On intelligence.
- A measure of the ability of a system to predict what happens next.
- Not measured by actual behavior or actions.
- Requires structured inputs: need for a patterned world.
- Prediction is the verification of reality and the essence of understanding.
- The odds that the same patterns occur randomly are low.
- The repeated pattern must therefore exist (at least at our level of observation).
- Learning combines bottom-up classification and top-down pattern prediction.
- Memory and learning are meaningless if inputs are random, without patterns.
- Parts of our outside world need to have (repeating) structure to be predictable.
- Used by all living systems to exploit the structure of the world.
Data comes in through the senses, flows into the cortex and, regardless of the nature of the pattern (speech, vision, somatosensory), is processed by the same cortical algorithm: the memory-prediction framework.
Memory – the cortex:
- stores sequences of patterns.
- recalls patterns auto-associatively.
- stores patterns in an invariant (essence, not details, reduced dimensionality) form.
- stores patterns in a hierarchy.
A memory system that is fed an ongoing stream of sensory inputs yields the potential for intelligent behavior.
Intelligence is measured by the capacity to remember and then predict patterns in the world: receive input from the senses, compare/combine that input with memory, then predict what happens next.
Prediction is the verification of reality. If you can reliably predict patterns through input from the senses over time (for instance, you continue to see or touch something, thereby confirming the continued presence of an object), the odds that those input patterns occurred in the same pattern over and over again randomly are low, so the object/pattern must really exist (at least at our level of observation). Prediction is the essence of understanding.
Learning is the combination of bottom-up classification of sensory inputs and top-down pattern/sequence prediction.
Memory and prediction are used by all living things (in varying degrees of sophistication) to exploit the structure of the world (for the benefit of reproduction).
The increase in human intelligence (compared to other animals) was driven by an increase in the cortex (the expanding memory capacity allowed for understanding more complicated patterns) and the cortex taking over most of the motor control areas (previously motor control was located in the old, reptilian brain; motor control in the cortex allows for more sophisticated predictions).
Machine intelligence can be achieved by feeding (any type of) sensory input streams into hierarchical memory models, allowing the system to store, learn and then predict patterns.