Predicting cortical oscillations with bidirectional LSTM network: a simulation study

It has been stated that up-down-state (UDS) cortical oscillation levels between excitatory and inhibitory neurons play a fundamental role in brain network construction. Predicting the time series behaviors of neurons in periodic and chaotic regimes can help in improving diseases, higher-order human activities, and memory consolidation. Predicting the time series is usually done by machine learning methods. In paper, the deep bidirectional long short-term memory (DBLSTM) network is employed to predict the time evolution of regular, large-scale UDS oscillations produced by a previously developed neocortical network model. In noisy time-series prediction tasks, we compared the DBLSTM performance with two other variants of deep LSTM networks: standard LSTM, LSTM projected, and gated recurrent unit (GRU) cells. We also applied the classic seasonal autoregressive integrated moving average (SARIMA) time-series prediction method as an additional baseline. The results are justified through qualitative resemblance between the bifurcation diagrams of the actual and predicted outputs and quantitative error analyses of the network performance. The results of extensive simulations showed that the DBLSTM network provides accurate short and long-term predictions in both periodic and chaotic behavioral regimes and offers robust solutions in the presence of the corruption process.