Harnessing adaptive dynamics in neuro-memristive nanowire networks for transfer learning
Nanowire networks (NWNs) represent a unique hardware platform for neuromorphic information processing. In addition to exhibiting synapse-like resistive switching memory at their cross-point junctions, their self-assembly confers a neural network-like topology to their electrical circuitry, something that is impossible to achieve through conventional top-down fabrication approaches. In addition to their low power requirements, cost effectiveness and efficient interconnects, neuromorphic NWNs are also fault-tolerant and self-healing. These highly attractive properties can be largely attributed to their complex network connectivity, which enables a rich repertoire of adaptive nonlinear dynamics, including edge-of-chaos criticality. Here, we show how the adaptive dynamics intrinsic to neuromorphic NWNs can be harnessed to achieve transfer learning. We demonstrate this through simulations of a reservoir computing implementation in which NWNs perform the well-known benchmarking task of Mackey-Glass (MG) signal forecasting. First we show how NWNs can predict MG signals with arbitrary degrees of unpredictability (i.e. chaos). We then show that NWNs pre-exposed to a MG signal perform better in forecasting than NWNs without prior experience of an MG signal. This type of transfer learning is enabled by the network's collective memory of previous states. Overall, their adaptive signal processing capabilities make neuromorphic NWNs promising candidates for emerging real-time applications in IoT devices in particular, at the far edge.