Adaptive behavior with stable synapses

Behavioral changes in animals and humans, triggered by errors or verbal instructions, can occur extremely rapidly. While learning theories typically attribute improvements in performance to synaptic plasticity, recent findings suggest that such fast adaptations may instead result from dynamic reconfiguration of the networks involved without changes to synaptic weights. Recently, similar capabilities have been observed in transformers, foundational architecture in machine learning widely used in applications such as natural language and image processing. Transformers are capable of in-context learning, the ability to adapt and acquire new information dynamically within the context of the task or environment they are currently engaged in, without changing their parameters. We argue that this property may stem from gain modulation-a feature widely observed in biological networks, such as pyramidal neurons through input segregation and dendritic amplification. We propose a constructive approach to induce in-context learning in an architecture composed of recurrent networks with gain modulation, demonstrating abilities inaccessible to standard networks. In particular, we show that, such architecture can dynamically implement standard gradient-based by encoding weight changes in the activity of another network. We argue that, while these algorithms are traditionally associated with synaptic plasticity, their reliance on non-local terms suggests that they may be more naturally realized in the brain at the level of neural circuits. We demonstrate that we can extend our approach to temporal tasks and reinforcement learning. We further validate our approach in a MuJoCo ant navigation task, showcasing a neuromorphic control paradigm via real-time network reconfiguration.