Font Size:
Sparse Signal Reconstruction via Kanerva's Sparse Distributed Memory
Last modified: 2021-10-18
Abstract
The phenomenon of memory has been studied by many neurobiologists. The variety of memory types includes short-term, long-term, sensory, topographic, semantic, immune, and so on. However, none of them works as a memory of the electronic devices: new information does not overwrite the old one, losing a small number of neurons does not affect the whole system, there is no explicit separation into address and data. Human memory is also capable of recognizing complex objects and structures without great efforts. The purpose of this article is to observe the representation of high-level objects with a set of features and relations in an artificial intelligence system. The specific sparse feature-based encoding is introduced. This representation is investigated subject to fast GPU implementation of a modified version of a well-known human memory model called Sparse Distributed Memory (SDM). A hybrid model of SDM and Compressed Sensing is proposed. Two techniques for reading sparse data from the hybrid model are designed and examined. Comparative analysis for both sparse and dense signal reconstruction problems is provided.
Full Text:
PDF