Pattern Machine

Amol Kelkar (kelkar.amol@gmail.com)

Abstract

Pattern Machine is a novel neural network architecture that represents hierarchical spatiotemporal patterns, capable of implementing cognitive homeostatic agents, which would yield Sentient AI systems. Here are the core tenets -

• Disentangle probability and uncertainty in representations (think predicting cat vs dog in daylight vs in dim twilight)
• Predictive processing: Use precision weighted mixing of bottom up signals and top-down predictions to yield oscillating attention toggling between inputs and predictions
• Self organized maps such that moving along a manifold represents a memory engram (episodic memory fragment). In other words, SOM neighborhood should represent state transition function.
• Each layer represents a memory bank of episodic memory fragments
• Locally predicted rewards that track global rewards that represent utility for homeostasis
• Local learning to preferentially remember rewarding episodes

Experiments

We have attempted more than a dozen different ways to implement Pattern Machine. Those experiments are in GitHub. Currently, we are back to exploring an architecture based on Spiking Neurons. It seems that evolution ended up at a fairly optimal solution to satisfy not only biological but also computational constraints and requirements!