Bio-inspired Deep Learning Model for Object Recognition

This paper proposes a bio-inspired deep learning architecture for object recognition and classification. The image samples are subjected to a saliency-based pre-processing step suitable for scene analysis and feature derivation. This preprocessing step bears similarities with the primate visual system which also assembles a saliency map. Thereafter, the deep learning model which relies upon the Hierarchical Temporal Memories (HTM) notion is utilized to form the corresponding feature vector. The latter HTM architecture consists of a tree shaped hierarchy of computational nodes where all nodes perform an identical procedure. Concerning the node operation, it forms representative vectors in order to sufficiently describe the input space. Afterwards, the representative vectors are utilized in order to derive spatial groups. The samples are expressed according to their degree of similarity with these groups using the L-1-norm minimization. The proposed bio-inspired scheme is compared with other state-of-the-art algorithms yielding remarkable performance.