PAIR-WISE TEMPORAL POOLING METHOD FOR RAPID TRAINING OF THE HTM NETWORKS USED IN COMPUTER VISION APPLICATIONS

摘要

In the paper, several modifications to the conventional learning algorithms of the Hierarchical Temporal Memory (HTM) - a biologically inspired large-scale model of the neocortex by Numenta - have been proposed. Firstly, an alternative spatial pooling method has been introduced, which makes use of a random pattern generator exploiting the Metropolis-Hastings algorithm. The original inference algorithm by Numenta has been reformulated, in order to reduce a number of tunable parameters and to optimize its computational efficiency. The main contribution of the paper consists in the proposal of a novel temporal pooling method -the pair-wise explorer - which allows faster and more reliable training of the HTM networks using data without inherent temporal information (e.g., static images). While the conventional temporal pooler trains the HTM network on a finite segment of the smooth Brownian-like random walk across the training images, the proposed method performs training by means of the pairs of patterns randomly sampled (in a special manner) from a virtually infinite smooth random walk. We have conducted a set of experiments with the single-layer HTM network applied to the position, scale, and rotation-invariant recognition of geometric objects. The obtained results provide a clear evidence that the pair-wise method yields significantly faster convergence to the theoretical maximum of the classification accuracy with respect to both the length of the training sequence (defined by the maximum allowed number of updates of the time adjacency matrix - TAM) and the number of training patterns. The advantage of the proposed explorer manifested itself mostly in the lower range of TAM updates where it caused up to 10 % relative accuracy improvement over the conventional method. Therefore we suggest to use the pair-wise explorer, instead of the smooth explorer, always when the HTM network is trained on a set of static images, especially when the exhaustive training is impossible due to the complexity of the given task.