Biometric System Laboratory Pattern Recognition by Hierarchical Temporal Memory

Hierarchical Temporal Memory (HTM) is still largely unknown by the pattern recognition community and only a few studies have been published in the scientific literature. This paper reviews HTM architecture and related learning algorithms by using formal notation and pseudocode description. Novel approaches are then proposed to encode coincidence-group membership (fuzzy grouping) and to derive temporal groups (maxstab temporal clustering). Systematic experiments on three line-drawing datasets have been carried out to better understand HTM peculiarities and to extensively compare it against other well-know pattern recognition approaches. Our results prove the effectiveness of the new algorithms introduced and that HTM, even if still in its infancy, compares favorably with other existing technologies. IERARCHICAL temporal memory (HTM) is a biologically-inspired computational framework recently proposed by Hawkins and George [1-3] as a first practical implementation of the memory-prediction theory of brain function presented by Hawkins in [4]. A private company, called Numenta 1 [5], was setup to develop HTM technology and to make available to researches and practitioners a complete development platform. A number of technical reports and presentations are available in Numenta website [5] to describe HTM technology, application and results, but at today few independent studies [6-12] have been published to validate this computational framework and to frame it into the state-of-the-art. HTM substantially differs from traditional neural network implementations (e.g., a multilayer perceptron) and can be conveniently framed into Deep Architectures [13][14]. In particular, Ranzato et al. [15] introduced the term Multistage Hubel-Wiesel Architectures (MHWA) to denote a specific subfamily of Deep Architectures. An MHWA is organized in alternating layers of feature detectors (reminiscent of Hubel and Wiesel’s simple cells) and local pooling/subsampling of features (reminiscent of Hubel and Wiesel’s complex cells); a final layer trained in supervised mode performs the classification. Neocognitron [16], Convolutional Networks [17][18], HMAX and its evolutions [19][20] are the best known implementations of MHWA. In analogy with MHWA, HTM alternates feature detection and feature pooling; however, in HTM feature pooling heavily relies on the temporal analysis of pattern sequences while in Neocognitron is hardwired and in Convolutional Network and HMAX is performed through simple spatial operators such as max or average. The temporal analysis and the modeling as a Bayesian Network make HTM similar in some aspects to Hierarchical [21] or Layered [22] versions of Hidden Markov Models (HMM); however, while HMM attempts to model the intrinsic temporal structure of input patterns 2 , HTM exploits time continuity (mainly during learning) for unsupervised derivation of invariant representations, independently of the static or dynamic nature of the input patterns. 1 the author of this paper has no business relationships with Numenta or with its founders, and has no commercial interest in promoting HTM technology. 2 in fact, the most successful HMM applications are in domains where patterns have an intrinsic temporal structure (e.g., speech recognition) or a spatial structure that can be naturally decomposed in subsequent parts (e.g., handwriting recognition). H