Autocorrelations in homeostatic spiking neural networks as a result of emergent bistable activity

Using a neuromorphic processor, we emulate networks of excitatory and inhibitory leaky integrate and fire neurons with spiking rates regulated by homeostatic plasticity. The latter incorporates stochastic updates that give rise to heterogeneous weight distributions. As predicted by theory, the network becomes more recurrent for decreasing input strength, which manifests in an increase of the autocorrelation time. Surprisingly, this rise can be attributed to emergent bistable population activity that (i) can be well approximated by a hidden Markov model, (ii) does not appear to vanish for increasing system sizes, and (iii) is likely stabilized by the heterogeneous weight distribution. In addition, we show that networks with bistable population activity allow for a more precise, yet slower representation of additional input that may still be read out once the input is removed.