Hierarchical Temporal Memory on the Automata Processor

While modern machine learning has made significant strides toward achieving high aptitude in various cognitive tasks, it does not provide any of the general abilities of mammalian intelligence. With a fundamentally different approach, Hierarchical Temporal Memory (HTM) is based on biological evidence that a common set of principles in the neocortex provides a diverse set of intelligent functions. Hierarchical temporal memories (HTMs) are biomimetic algorithms that can similarly be trained to perform inference and prediction on any temporal datastream. The Automata Processor (AP) is a configurable silicon implementation of nondeterministic finite automata, designed for massively parallel pattern matching. Key correspondences between counter-extended nondeterministic finite automata and the HTM activation model indicate use of the AP as an efficient hardware accelerator. In this article, the authors introduce a methodology for synthesizing HTMs on the Automata Processor, demonstrate three prediction applications on their model, and show its potential to achieve between 137 to 446 times speedup over the CPU.