A Macrocolumn Architecture Implemented with Temporal (Spiking) Neurons

07/11/2022

∙

by   James E. Smith, et al.

∙

0

∙

share

With the long-term goal of reverse-architecting the computational brain from the bottom up, the focus of this document is the macrocolumn abstraction layer. A basic macrocolumn architecture is developed by first describing its operation with a state machine model. Then state machine functions are implemented with spiking neurons that support temporal computation. The neuron model is based on active spiking dendrites and mirrors the Hawkins/Numenta neuron model. The architecture is demonstrated with a research benchmark in which an agent uses a macrocolumn to first learn and then navigate 2-d environments containing randomly placed features. Environments are represented in the macrocolumn as labeled directed graphs where edges connect features and labels indicate the relative displacements between them.