Researchers Develop More Energy-Efficient Analog Neural Networks Through Novel Training Algorithm
• Researchers develop algorithm to efficiently train analog neural networks, providing energy-efficient alternative to digital networks
• Method aligns more closely with human learning by updating physical system layers locally rather than using backpropagation
• Algorithm tested successfully on experimental acoustic, optical, and microwave systems to classify data
• Approach aims to increase scalability to larger, deeper physical neural networks in the future
• Analog systems enabled by new training method offer promise for reducing environmental impact of deep learning