AI Hardware Design Could Filter Out Irrelevant Visual Data to Reduce Energy Use
NUS researchers built a molybdenum disulfide transistor that can act as a logic filter, a neuron or a synapse, and used it to block irrelevant visual data before a spiking neural network processes it. In simulations, cropping digits cut active hardware tiles by about 36% for a 0.3-point accuracy loss, and a traffic scene used about 59% less energy for vehicle recognition and 77% less for licence plates. We explain how it works, check the arithmetic and list what the press release leaves out.
