Improving full-FORCE with dynamical data coupling and multilayer architecture

Yue Yin, Emre Neftci

Abstract: Training Deep Recurrent Neural Networks (Deep- RNNs) using Back Propagation Through Time (BPTT) has shown tremendous success in improving benchmark performance and solving real-world problems. However, the non-locality of loss functions for deep networks and the requirement of parallel computing hardware such as GPUs to propel learning make mapping gradient-based RNNs onto neuromorphic devices challenging. As a result, gradient-free alternatives like reservoir computing have emerged as more realistic and bio-plausible options to train RNNs on neuromorphic substrates. This study improves the stability of a target-based method called full-FORCE for training RNNs by dynamically coupling targets with the network. We show that this coupling has the same stabilizing effect as when learning online. We also extend full-FORCE to multiple layers while permitting online local weight updates for each layer. The proposed network outperforms the original version on tasks including pattern generation and interval matching. It can also classify eye states using electroencephalogram (EEG) recordings with higher accuracy than full-FORCE (0.886±0.030 versus 0.772±0.052), showing potential in biomedical applications such as real-time computation on brain-interfacing devices.