NEURITHMIC SYSTEMS LLC

Developing extremely efficient, neuromorphic, single/few-trial learning and probabilistic inference based on a unique sparse distributed representation (SDR) format for representing information. The model, Sparsey, uses a revolutionary learning algorithm that preserves spatiotemporal similarity from the input space (e.g., of videos) to the representation space (SDRs), without requiring new inputs to be explicitly compared to previously stored inputs. The memory traces of spatiotemporal inputs (sequences) are mapped to chains of SDRs and the spatiotemporal similarity structure over the sequences automatically emerges, in the pattern of intersections over the SDRs, as a by-product of storing specific memory traces. Thus, the statistical (similarity) structure over the inputs, a.k.a. generative model, semantic memory, resides in physical superposition with the episodic memory traces of the individually experienced items. Sparsey is generic, scalable, and super-efficient technology that can be applied to any manner of multivariate time series, including multimodal multivariate time series. By 'super-efficient' we mean that the time it takes to: a) store a new input: b) retrieve the best-matching (or, most relevant, most likely stored input; and in fact c) retrieve all stored inputs in descending similarity (relevance, likelihood) order; remain fixed as the number of stored items grows.