“What-Where” sparse distributed invariant representations of visual patterns

Abstract

Although modern deep learning approaches have achieved astounding results in most visual pattern recognition tasks, they do it using large datasets of labeled data. Besides the fact that, in many applications, such labels are costly to obtain, the need for them is not observed in a biologically intelligent machine like the human brain. “What-Where” sets were proposed as a way to represent visual patterns in a manipulatable manner, where two-dimensional geometric transformations can be exploited to increase invariance, and thus reduce the need for large amounts of training data. However, the cornerstone of classification using these sets is a similarity measure that implicates a time-consuming computation due to the unstructured nature of sets. In this work, we propose a grid-based coding strategy to represent the sets as sparse binary vectors. By doing so, we achieve three main advantages: first, leveraging pointer-coding of active bits, we reduce the time complexity of the similarity computation from quadratic to linear in the number of elements of the smaller set being compared; second, we use the theoretical framework of sparse representations to justify the classification robustness exhibited in the original work; third, we bring the model under the widely accepted biological constraint that populations of neurons in the brain code sparse representations.