RT Journal Article
SR Electronic
T1 Biologically Plausible Learning in Recurrent Neural Networks for Flexible Decision Tasks
JF bioRxiv
FD Cold Spring Harbor Laboratory
SP 057729
DO 10.1101/057729
A1 Thomas Miconi
YR 2016
UL http://biorxiv.org/content/early/2016/06/07/057729.abstract
AB Recurrent neural networks operating in the near-chaotic regime exhibit complex dynamics, reminiscent of neural activity in higher cortical areas. As a result, these networks have been proposed as models of cortical computation during cognitive tasks. However, existing methods for training the connectivity of these networks are either biologically implausible, and/or require an instantaneous, real-time continuous error signal to guide the learning process. The lack of plausible learning method may restrict the applicability of recurrent neural networks as models of cortical computation. Here we introduce a biologically plausible learning rule that can train such recurrent networks, guided solely by delayed, phasic rewards at the end of each trial. We use this method to learn various tasks from the experimental literature, showing that this learning rule can successfully implement flexible associations, memory maintenance, nonlinear mixed selectivities, and coordination among multiple outputs. The trained networks exhibit complex dynamics previously observed in animal cortex, such as dynamic encoding and maintenance of task features, switching from stimulus-specific to response-specific representations, and selective integration of relevant input streams. We conclude that recurrent neural networks can offer a plausible model of cortical dynamics during both learning and performance of flexible behavior.