Rapidly reconfigurable dynamic computing in neural networks with fixed synaptic connectivity

Learning and memory in the brain’s neocortex have long been hypothesised to be primarily mediated by synaptic plasticity. Extensive research in artificial neural networks has shown that training networks by adjusting connection weights faces computational challenges, including large parameter spaces and the tendency of new learning to interfere with previous learning (catastrophic forgetting). We propose that the brain, which is resistant to these challenges, can also learn by modulating the excitability of each neuron in a network rather than changing synaptic strengths. We show here that learning a task-specific set of bias currents enables a feedforward or recurrent network with fixed and randomly assigned connections to perform well on and switch between dozens of tasks, including regression, classification, autonomous time series generation, a game and motor control. Bias-only learning also provides a mechanistic account of how representational drift and structured neuron-level variability can coexist with stable population-level computation. The authors test whether changing neuronal excitability can train feedforward and recurrent neural networks while all synaptic connections remain fixed. This mechanism supports diverse dynamical, classification, regression and control tasks, and preserves stable outputs despite changes in individual-neuron activity.