Sound Waves Give Neuromorphic Chips an Edge in Mimicking the Brain
Университет Аризоны
A new study suggests that using sound waves in neuromorphic devices can better mimic biological neurons, operating faster and with greater energy efficiency than electronic counterparts. The acoustic synapse developed by researchers uses multiple phi-bits to perform simultaneous computations, achieving 96.7% accuracy in flower classification while consuming at most one-tenth the power of state-of-the-art electronic neuromorphic hardware.
Researchers at the University of Arizona, led by Xiaodong Yan, have developed an acoustic synapse using sound waves that can better mimic biological neurons. The device consists of three aluminum rods connected by epoxy glue, with ultrasonic transmitters and sensors attached. Using sound waves, the researchers encoded data and modulated the phase of phi-bits to mimic synaptic plasticity, allowing the device to strengthen or weaken connections over time. In tests, the acoustic synapse outperformed a conventional multilayer perceptron neural network in classifying iris flowers, achieving 96.7% accuracy with only 39 parameters and reaching peak accuracy 20% faster. The device consumes at most one-tenth the power of current electronic neuromorphic hardware. Additionally, by adding an extra rod, the system could mimic neuromodulatory processes such as those involving dopamine and serotonin, enabling flexible adaptation similar to biological brains. The findings were published in Science Advances on June 12.
- Сокращения
- MLP = Multilayer Perceptron — многослойный перцептрон
Source: IEEE Spectrum AI —
original