Dendroplex: Synthesis, Simulation, and Validation of Hierarchical Temporal Memory on the Automata Processor
Hierarchical temporal memory is a growing paradigm for machine learning with an approach to achieving true machine intelligence based on the algorithmic properties of the neocortex. The Automata Processor is a non-von Neumann MISD processor that is a silicon implementation of non-deterministic finite automata. Many natural correspondences between the execution models of the AP and HTM that make the AP a good choice for accelerating HTM. In this work, we present a model for dendritic activation processing implemented in the Automata Processor, a synthesis methodology for construction of general HTM models, and the tools developed to facilitate this workflow. We show the effectiveness of this approach for acceleration of HTM Spatial Pooler and Temporal Memory functions by up to 446X over the baseline CPU software implementation.