New weightless neural networks could slash AI energy use
University of Texas researcher Lizy K. John is developing weightless neural networks that use lookup tables instead of multiplication, potentially slashing AI energy use and model sizes.
Lizy K. John, a professor at the University of Texas at Austin, is pioneering weightless neural networks that replace traditional, energy-intensive multiplication with interconnected lookup tables. By bypassing the mathematical operations tracing back to the 1943 McCulloch-Pitts neuron model, these networks can be under a thousandth of the size or 1,000 times faster than conventional alternatives. For instance, in a medical monitoring test, John’s team reduced a model's size from a standard 17 megabytes down to just 14 kilobytes. In keyword-spotting tasks, where standard industry models consume over 5,000 nanojoules per inference, this weightless architecture requires a mere 42 to 79 nanojoules.
This drastic reduction in computational overhead allows these models to run on highly constrained hardware. The researchers successfully deployed an arrhythmia detector on a flexible plastic substrate containing only about 10,000 logic gates, a stark contrast to the billions of gates found on modern 2-nanometer or 5-nanometer silicon chips. Because FPGAs have featured built-in lookup tables for over 30 years, these weightless networks can run efficiently on single FPGAs without requiring expensive, power-hungry GPUs for inference.
For AI practitioners, this paradigm shift enables true edge computing on battery-powered sensors, such as medical patches monitoring ECG, EEG, or blood pressure. Instead of constantly transmitting raw data every millisecond to a central server, devices can process information locally, enhancing both energy efficiency and user privacy. While the technology has primarily targeted small-scale sensor tasks, John's team has already replaced the multilayer-perceptron portion of transformer networks, which constitutes roughly half of the model. If researchers can successfully adapt the attention layers, weightless architectures could eventually scale to large language models, offering a path away from the massive energy footprint of modern generative AI.
This is our own summary of reporting by IEEE Spectrum AI
