Hierarchical temporal memory : an investigative look into a new cortical algorithm -

Abstract

The connectionist approach to artificial intelligence has made many attempts to capture the essence of intelligence by adopting some of the principles underlying the operation of the human brain. Since its emergence in the 1950s, this field has been evolving with novel algorithms finding their way to the forefront. One such algorithm is Hierarchical Temporal Memory (HTM). While lacking in maturity, HTM’s approach to prediction and online learning has garnered the interest of machine learning practitioners. Major players in the tech industry like IBM have also taken notice of HTM as of April, 2015 and have dedicated their resources to investigating its merit. The work presented in this thesis reflects that interest and seeks to establish where HTM stands in the grander machine learning scheme. To that end, we take advantage of the modular design of HTM and propose a number of tests aiming to evaluate its individual modules as well as HTM as a whole. In modular testing, HTM is put through a series of basic tests for clustering and sequence learning tasks. These experiments reveal some of the spatial and temporal pooling promise of HTM but ultimately expose some of its inherent flaws. More involved experiments utilizing the entire HTM stack go on to demonstrate that further with HTM struggling when adapted to multi-class classification tasks and anomaly detection on streaming data.