Energy-efficient AI training with memristors

Novel probabilistic training techniques for memristor hardware promise dramatic energy savings and improved accuracy, signalling a major step towards practical and sustainable analog AI systems.
Energy-hungry AI training could be transformed as memristor-based systems adopt probabilistic updates that cut power use while extending hardware lifespan.

Scientists in China developed an error-aware probabilistic update (EaPU) to improve neural network training on memristor hardware. The method tackles accuracy and stability limits in analog computing.

Training inefficiency caused by noisy weight updates has slowed progress beyond inference tasks. EaPU applies probabilistic, threshold-based updates that preserve learning and sharply reduce write operations.

Experiments and simulations show major gains in energy efficiency, accuracy and device lifespan across vision models. Results suggest broader potential for sustainable AI training using emerging memory technologies.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!