Optimizing Hierarchical Temporal Memory for Multivariable Time Series

Hierarchical Temporal Memory (HTM) is an emerging computational paradigm consisting of a hierarchically connected network of nodes. The hierarchy models a key design principle of neocortical organization. Nodes throughout the hierarchy encode information by means of clustering spatial instances within their receptive fields according to temporal proximity. Literature shows HTMs' robust performance on traditional machine learning tasks such as image recognition. Problems involving multi-variable time series where instances unfold over time with no complete spatial representation at any point in time have proven trickier for HTMs. We have extended the traditional HTMs' principles by means of a top node that stores and aligns sequences of input patterns representing the spatio-temporal structure of instances to be learned. This extended HTM network improves performance with respect to traditional HTMs in machine learning tasks whose input instances unfold over time.