Measuring the Storing Capacity of Hyperdimensional Binary Vectors

Abstract

Hyperdimensional computing is a model of computation based on the properties of high-dimensional vectors. It combines characteristics from artificial neural networks and symbolic computing. An area where hyperdimensional computing can be applied is natural language processing, where vector representations are already present in the form of word embedding models. However, hyperdimensional computing encodes information differently, its representations can include the distributional information of a word in a given context and it can also account for its semantic features. In this work, we investigate the storing capacity of hyperdimensional binary vectors. We present two different configurations in which semantic features can be encoded and measure how many can be stored, and later retrieved, within a single vector. The results presented in this work lay the foundation to develop a concept representation model with hyperdimensional computation.