Sound Waves Enable Neuromorphic Chips to Mimic Brain Function More Efficiently

Researchers at the University of Arizona have developed a novel neuromorphic computing device that utilizes sound waves to mimic biological neurons, offering significant improvements in speed and energy efficiency over conventional electronic AI chips. This new approach aims to overcome the limitations of current neuromorphic hardware, which, despite being more energy-efficient than traditional chips, still possess a fraction of the connectivity found in the human brain. The device, described in a study published in Science Advances, uses acoustic waves to encode data, enabling multiple computations simultaneously within a single unit, thus reducing the need for extensive wiring and hardware complexity. Xiaodong Yan, an assistant professor at the University of Arizona and lead author, explained that this could lead to more compact, parallel, and efficient hardware for tasks like pattern recognition and data analysis. The system employs acoustic synapses that can modulate the phase of sound waves, similar to how biological synapses strengthen or weaken over time, a property known as synaptic plasticity. This allows the device to learn and adapt, mimicking memory formation and fading. In experiments, the acoustic synapse demonstrated superior performance in classifying iris flowers compared to a conventional multilayer perceptron (MLP), achieving higher accuracy with fewer parameters and reaching peak performance 20 percent faster. The researchers estimate the new device consumes up to ten times less power than current state-of-the-art electronic neuromorphic hardware. Furthermore, the acoustic synapse can mimic the effects of neuromodulators like dopamine and serotonin, which influence synaptic strength and learning in the brain, allowing for flexible adaptation to different cognitive states. This flexibility could enable smaller, adaptable neural networks capable of performing diverse functions without requiring entirely separate architectures for each task, a significant advancement over current systems.