AI Models Working Harder Than They Need To? Weightless Neural Networks Offer a Lookup-Based Alternative

Professor Lizy K. John of the University of Texas at Austin argues that current AI models, which rely heavily on multiplication operations, may be doing more work than necessary. For five years, she has worked on weightless neural networks that use interconnected lookup tables instead of multiplying inputs by weights, reducing computation to consulting stored binary answers. On tasks like human-activity recognition, ECG monitoring, and keyword spotting, her team's networks achieve 1,000 times smaller size and energy use while maintaining accuracy. For example, their keyword spotting model uses only 42 to 79 nanojoules per inference compared to over 5,000 for the current best industry model. John has already replaced the multilayer-perceptron portion of transformer networks, about half of a chatbot's architecture, and aims to replace the attention mechanism as well. She believes weightless networks could enable AI on tiny, low-power devices like bendable plastic substrates for medical sensors. Training still occurs on GPUs, but John hopes to eventually use FPGA-based hardware due to the natural fit of FPGA's lookup tables.