NB-CNS Seminar - Shih-Chii Liu

"Brain-inspired technology for neuromorphic edge AI and visual-assistive neuroprosthesis"

This talk discusses the use of brain-inspired technology in two areas: 1) for neuromorphic audio edge AI and 2) for a brain-assistive neuroprosthesis. The first part will present a neuro-inspired leveraging of cochlea auditory processing and brain-like dynamic sparsity to boost the dynamic range and energy efficiency of edge hardware AI. We show how the dynamic sparsity inspired by spikes can be practically embedded into recurrent networks by adopting a delta neuron model. This model is built into a mixed-signal 75 dB input DR, 2mm2, 65nm CMOS audio System-on-Chip demonstrating end-to-end user intent understanding spoken language understanding with <10uW power; and won the 2025 ISSCC best demo award. The second part will discuss ongoing work in algorithms and hardware for a visual-assistive cortical neuroprosthesis. It will first introduce a data-driven neural control method using a deep learning framework to generate optimal multi-electrode stimulation patterns that evoke targeted neural responses. Using data from a 96-channel Utah electrode array implanted in the occipital cortex of a blind participant, a deep neural network was trained to predict single-trial evoked responses. The network was subsequently used in two complementary control strategies: a learned inverse network for real-time stimulation synthesis and a gradient-based optimizer for precise targeting of desired neural responses, reliably evoking more stable percepts. It also introduces an assistive technology setup that allows comparative evaluation of prosthetic vision algorithms on a wearable VR system.

Faculty Host: Thanos Siapas