A Self-adaptive Hardware With Resistive Switching Synapses For Experience-based Neurocomputing

This work proposes a neuromorphic hardware based on Silicon Oxide (SiOx) RRAM devices to join state-of-the-art accuracy and bio-inspired plasticity for autonomous and resilient navigation at low-power. The network relies on bio-inspired algorithms, such as STDP and plastic homeostasis, to adjust the parameters along a temporal sequence, as in recurrent neural networks (RNNs). The RRAM devices are used for both Hebbian learning processes and to map the recurrent internal state of each neuron. To test the resilience of the hardware, a two-dimensional dynamic maze showing environmental changes in time is experimentally configured in a field-programmable-gate-array (FPGA).