Modeling persistent temporal patterns in dissociated cortical cultures using reservoir computing

A model of a neural system should be able to reproduce the temporal patterns under the same input and/or initial state, which is a minimal requirement for a network-level model to reveal the information encoded in such patterns. Our approach taken here is to employ a generic model (a reservoir network) that displays a rich repertoire of complex spatiotemporal patterns to be matched with the observed biological patterns by parameter tuning. More specifically, we employ an Echo State Network (ESN) [2] with leaky integrator neurons as a modeling tool. Here, we consider cultures of dissociated cortical tissue recorded with microelectrode arrays (MEA) as an example of biological neural networks without specific connectivity and simulate the corresponding burst types based on a cue signal. The cue signal is composed of a snapshot (10 ms) of the individual firing rates recorded at each electrode at burst onset and serves as an indicator of the current dynamic state of the network. A simple readout training of the ESN yields a predictive model of the temporal activity pattern in the global firing rate. The simulated pattern displays a high correlation with the actual one observed in the culture (Figure 1). The model can also be used to visualize the underlying structure in the recorded signals.