Sound Waves Give Neuromorphic Chips a Brain-Simulating Edge

A new acoustic device that processes information with sound waves can mimic biological synapses more closely than electronic equivalents, operating faster and using at most one-tenth the power of state-of-the-art neuromorphic hardware, according to a study published 12 June in Science Advances [1]. The device, built from three aluminum rods each roughly 60 centimeters long and 1.25 centimeters wide, uses ultrasonic transmitters and sensors to encode data into phase bits, or phi-bits, that coexist in a single space [1]. Unlike conventional bits that each represent only a 0 or 1 and require separate physical components, phi-bits can represent multiple variables simultaneously, enabling parallel computation through classical analogues of quantum logic gates [1]. Most neuromorphic devices today function essentially as a single artificial synapse, whereas a biological neuron may possess thousands of synapses [1]. Artificial neural networks, which loosely model the brain's structure, consist of connected nodes that process signals through weighted edges analogous to synapses [2]. The acoustic approach allows multiple simultaneous computations without the wiring complexity and energy cost of linking many separate electronic devices [1]. In a flower-classification experiment involving 150 iris specimens, the acoustic synapse coupled with three digital neurons achieved 96.7 percent accuracy using only 39 parameters and reached peak accuracy 20 percent faster than a conventional multilayer perceptron [1]. An MLP required nine neurons and more parameters to reach comparable performance [1]. "In a topological acoustic synapse, the acoustic wave interactions help transform and organize information before the final readout," said Xiaodong Yan, an assistant professor at the University of Arizona [1]. The emerging field of topological acoustics exploits previously unknown properties of sound waves to create circuits where acoustic energy flows with virtually no dissipation [1]. The researchers also demonstrated that adding an extra rod allowed the system to mimic neuromodulators such as dopamine or serotonin, which adjust synaptic sensitivity, speed, and learning strength [1]. A single biological synapse may be influenced by as many as 10 neuromodulators simultaneously, but replicating this in conventional hardware typically demands far more complex designs [1]. "Neuromodulators let the brain use one circuit to perform different functions depending on the context," said Brad Aimone, a researcher at Sandia National Laboratories in Albuquerque, who was not involved in the study [1]. "So instead of an enormous neural network, you could have smaller neural networks that can use the equivalent of neuromodulators to adjust themselves for whatever's going on." [1] Sandia National Laboratories, located in Albuquerque, is one of three laboratories of the National Nuclear Security Administration and a major federal research presence in New Mexico [10]. The city has been a scientific and military hub since World War II [11].