Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing (arxiv.org)

Recent advancements in physical neural computing have highlighted a shift away from traditional silicon-based hardware, exploring diverse substrates like memristive devices, photonic circuits, mechanical metamaterials, and even living neural tissue. This evolution is significant as silicon's constraints in energy consumption and data movement become more pronounced. The push towards physical neural computation is particularly vital for the deployment of on-device and edge AI, where closer integration of computation, sensing, and memory can vastly enhance efficiency in resource-limited contexts. A comprehensive survey of these emerging technologies aims to unify the field by mapping neural primitives to the unique mechanisms of various substrates, while addressing key engineering challenges such as scalability, precision, and interfacing. Additionally, the introduction of a benchmarking scheme offers a standardized way to evaluate these platforms against specific tasks, revealing that no single substrate excels universally. Instead, various physical neural systems complement each other across diverse applications, ranging from ultrafast signal processing to biochemical decision-making, indicating a promising future for the AI/ML landscape.