Brain Inspired One Shot Learning Method for HD Computing

Abstract

The brain’s power of perception, learning, thought, language, concepts, adaptivity, ability to tolerate noise etc are amazing to the human kind for a long time and it is a big challenge for us to model a platform with these properties. In this work, we propose a new architecture for applications in high dimensional computing. The brain-inspired high dimensional computing architecture involves generation and storage of large strings of random bits using hypervectors. A True Random Number Generator (TRNG) using Fibonacci and Galois ring oscillator is for generating high dimensional vectors. Digital post processing of random data using Linear Feedback Shift Registers (LFSR) is also introduced. High dimensional binary vectors are then used for representing entities. High degree of parallelism, noise tolerance, learning by analogy are some of the characteristics that make it possible for use in brain inspired computing. Operations like superposition and binding are defined for High dimensional vectors. The algebra of high dimensional vectors make this architecture suitable for certain cognitive tasks such as learning by analogy, learning for classification, etc. The proposed high dimensional computing architecture can be used for language is implemented using 2-D architecture for High Dimensional (HD) computing and validated against the benchmark datasets. This architecture introduces a new method of learning known as one shot learning, considering the effectiveness in resource consumption and computation flexibility. The power consumption and related design issues are presented in this work. The experiments carried out for the language identification problem are motivating and opens up a new paradigm for machine learning and classification tasks such as cancer detection using benign and malignant cells, spam mail detection and so on.