Nonlinear Computing in Silicon A large part of current AI hardware focuses on accelerating linear operations, in particular matrix-vector multiplication. Yet the expressive power of neural networks critically depends on nonlinear transformations. This motivates a complementary approach in which nonlinear computation is performed directly by the physics of devices. I will introduce reconfigurable nonlinear-processing units (RNPUs), silicon-based physical computing primitives that combine strong intrinsic nonlinearity with electrical reconfigurability. RNPUs have been trained for classification tasks, optimized directly in materia, and applied to temporal signal processing and analogue speech recognition [1-4]. More recently, we have shown how such devices can implement the learnable nonlinear functions of Kolmogorov-Arnold Networks directly in hardware [5], while complementary device studies identify space-charge effects as the physical origin of their tunable nonlinear response [6]. Together, these results illustrate how computation can increasingly be embedded in the physical properties of materials and devices rather than being imposed entirely through conventional digital architectures. This concept is potentially interesting for neuroscience, where sensing, computation and actuation ultimately interface with an analogue physical system: the nervous system itself. Physical and analogue computing could enable local, low-power processing close to neural interfaces, reducing data movement and potentially allowing implants to become more adaptive and autonomous. I will briefly discuss how these ideas connect to our work on next-generation neural interfaces, including high-performance Pt/PtOx electrode materials for recording and stimulation and ongoing efforts toward flexible, long-term implantable systems [7]. More broadly, this points toward neural implants in which sensing, processing and stimulation are increasingly integrated into the physical interface itself.
Organizer
Zhenyu Gao
z.gao@erasmusmc.nl