Reconfigurable silicon photonics extreme learning machine with random non-linearities as neural processor and physical unclonable function

Abstract

We propose RN (reservoir computing)-ELM, an extreme learning machine in which the conventional architecture—fixed random linear synapses followed by a uniform activation function—is replaced by an array of fixed random non-linear synapses. This shift is motivated by recent edge-non-linearity neural architectures (Kolmogorov–Arnold and polynomial-regression networks), in which heterogeneity of the non-linearities yields higher expressivity within shallower networks. RN-ELM is realized as a hybrid analog–digital engine: the physical layer is a silicon-photonic bank of optical filters whose passive frequency-to-power response supplies the heterogeneous non-linearities, and the digital layer is a single linear regression. The non-linearities are intrinsically power-independent and require no active reconfiguration. The architecture is evaluated using an experimentally measured transfer function of an all-pass micro-ring resonator implemented on a field-programmable photonic gate array, which serves as the kernel of simulations targeting a deployed silicon-on-insulator application specific integrated circuit. RN-ELM is assessed in two complementary roles. As a machine-learning processor on the Santa-Fe chaotic time-series benchmark, it reaches a normalized mean squared error (NMSR) of 0.053 with only five optical filters, on par with state-of-the-art photonic implementations that use two orders of magnitude more physical nodes. As a physical unclonable function, the same fabrication-induced waveguide imperfections that supply the random non-linearities double as the entropy source of an authentication key, with a cloning probability as low as 10−15. RN-ELM thus unifies compact photonic neural processing and intrinsic hardware authentication in a single passive substrate.