Fixed Frame Temporal Pooling

Abstract

Applications of unsupervised learning techniques to action recognition have proved highly competitive in comparison to supervised and hand-crafted approaches, despite not being designed to handle image processing problems. Many of these techniques are either based on biological models of cognition or have responses that correlate to those observed in biological systems. In this study we apply (for the first time) an adaptation of the latest hierarchical temporal memory (HTM) cortical learning algorithms (CLAs) to the problem of action recognition. These HTM algorithms are both unsupervised and represent one of the most complete high-level syntheses available of the current neuroscientific understanding of the functioning of neocortex. Specifically, we extend the latest HTM work on augmented spatial pooling, to produce a fixed frame temporal pooler (FFTP). This pooler is evaluated on the well-known KTH action recognition data set and in comparison with the best performing unsupervised learning algorithm for bag-of-features classification in the area: independent subspace analysis (ISA). Our results show FFTP comes within 2% of ISA's performance and outperforms other comparable techniques on this data set. We take these results to be promising, given the preliminary nature of the research and that the FFTP algorithm is only a partial implementation of the proposed HTM architecture.