New Learning System for Scalable and Energy-Efficient AI, Explain Researchers

Deep-learning models have served to be highly useful tools to make projections and to solve real-world problems that involve the analysis of data. Nonetheless, deep-learning models require extensive training in physical data centers, before they are deployed in real devices and software, which can be time and energy consuming.

In a new development, a joint effort of researchers at Texas A&M University, Sandia National Laboratories, and Rain Neuromorphics has led to devise of a new system that can train deep learning models more efficiently and on a larger scale.

The details of the system appear in a paper published in Nature Electronics. The paper explains how the system relies on the use of memristor crossbar hardware and new training algorithms that can execute multiple operations in one go.

“Meanwhile, AI is mostly associated with face recognition in smart phones, health monitoring in smart watches, etc., but most of AI, in terms of energy spent, entails training of AI to perform these tasks,” stated the senior author of the study.

Training happens in data centers of the size of warehouse, which is very costly economically as well as in terms of carbon footprint.

Essentially, the objective of the researchers was to formulate an approach that could reduce the financial costs and carbon footprint associated with training AI models, thus making the large-scale implementation more sustainable and easier. This required overcoming two major limitations of current AI training practices.

The first challenge is associated with the use of inefficient hardware frameworks based on graphics processing units, which are not inherently designed for deep learning models. The second challenge entails the use of ineffective and math-laden software tools.