Memristor-based learning systems provide energy-efficient and scalable AI training.

A memristor-based crossbar-based AI system that is scalable and energy efficient

Deep-learning models are highly useful for solving tasks that require the analysis of real data. These models are not without their benefits, but they require intensive training in data centers before they can be used in actual software or devices, such as mobile phones. This is both time-consuming and energy-consuming.

Researchers from Texas A&M University Rain Neuromorphics, and Sandia National Laboratories developed a new system to train deep learning models on a large scale and more efficiently. This system was introduced in Nature Electronics and relies on new training algorithms as well as memristor hardware that can perform multiple operations simultaneously.

The study’s senior author, Suhas Kumar, said that most people associate AI with smart watches and smart phones for face recognition, health monitoring, etc. But the majority of AI energy is spent on training AI models to do these tasks.