Physical Review Research (Jul 2023)
Autocorrelations from emergent bistability in homeostatic spiking neural networks on neuromorphic hardware
Abstract
A fruitful approach towards neuromorphic computing is to mimic mechanisms of the brain in physical devices, which has led to successful replication of neuronlike dynamics and learning in the past. However, there remains a large set of neural self-organization mechanisms whose role for neuromorphic computing has yet to be explored. One such mechanism is homeostatic plasticity, which has recently been proposed to play a key role in shaping network dynamics and correlations. Here, we study—from a statistical-physics point of view—the emergent collective dynamics in a homeostatically regulated neuromorphic device that emulates a network of excitatory and inhibitory leaky integrate-and-fire neurons. Importantly, homeostatic plasticity is only active during the training stage and results in a heterogeneous weight distribution that we fix during the analysis stage. We verify the theoretical prediction that reducing the external input in a homeostatically regulated neural network increases temporal correlations, measuring autocorrelation times exceeding 500ms, despite single-neuron timescales of only 20ms, both in experiments on neuromorphic hardware and in computer simulations. However, unlike theoretically predicted near-critical fluctuations, we find that temporal correlations can originate from an emergent bistability. We identify this bistability as a fluctuation-induced stochastic switching between metastable active and quiescent states in the vicinity of a nonequilibrium phase transition. Our results thereby constitute a complementary mechanism for emergent autocorrelations in networks of spiking neurons with implications for future developments in neuromorphic computing.