Memristive Dynamical Spiking Neural Networks with Spatiotemporal Heterogeneity

We propose a fully memristive spiking neural network (MSNN) that incorporates spatiotemporal heterogeneity to improve temporal representation and fault tolerance. In our proposed work, each neuron possesses a distinct time constant (spatial heterogeneity) that evolves over time in response to input stimuli (temporal heterogeneity), enabling diverse, adaptive, and temporally rich responses. Both synaptic and neuronal behaviors are modeled using SPICE-level analog memristors, and the network is trained end-to-end using backpropagation through time (BPTT) in a differentiable framework. This approach eliminates the need for digital interfacing circuits such as ADCs or explicit comparators, supporting compact and efficient hardware deployment. Evaluations on the MNIST and DVS128 Gesture datasets show competitive accuracy and significantly improved robustness to hardware faults, such as stuck-at errors in RRAM cells. These results demonstrate the effectiveness of spatiotemporally heterogeneous MSNNs for scalable, reliable neuromorphic computing.