These are my notes on the following excellent video, by Jeff Hawkins. For reference, I buy it. This works for me. However, it’s rather old and I just wish we were further along, so I could use it to do stuff.
Does anyone know of an updated summary of comparable clarity?
How to do machine intelligence
1. Discover operating principles of neocortex
2. Build systems based on these principles
The neocortex is a memory system, which receives a high velocity data stream and learns a model that allows it to:
• make predictions
• detect anomalies
• perform actions
The top 6 principles of neocortical function are:
1. On-line memory system: no batch processing
2. Hierarchy of memory regions
○ highly connected
○ all regions broadly perform the same functions
3. Sequence memory: 90% of memory is patterns over time
4. Sparse Distributed Representations
5. All regions are sensory and motor
Claim: these 6 principles are both necessary and sufficient for biological and machine intelligence.
• many bits (thousands)
• few 1’s, mostly 0’s (eg 2000 bits,2% active)
• each bit has semantic meaning
• meaning of bit learned not assigned
• two SDRs with shared bits have semantic similarity
• only need to store the active bits (and a subset is OK)
• can union SDRs prior to compare
Claim: SDRs are the future of intelligent machines, there is no other way.
• self-driving cars