Go to NeurIPS 2025 Workshop NeurReps homepageNeuromodulator-inspired task switching using a fixed weight-space transformation
Keywords: neuromodulator, switchable, neuroAI
TL;DR: We explore a mechanism for task switching in neural networks that is inspired by chemical neuromodulation.
Abstract: In biological systems, chemical neuromodulators provide a means to broadly reconfigure the circuit activity of large populations of neurons. This enables switching between distinct behavioral states with control of just a few non-local variables. Applying such control mechanisms to deep neural networks could enhance their flexibility. This approach examines how learning of switchable behavior can occur when the task transformation itself is constrained and networks are free to evolve around the transformation. We explore the implementation of chemical neuromodulator inspired control by demonstrating a network that can switch between AND, OR and XOR gates by applying a single randomly initialized affine transformation directly to its weights. We train this switchable network by co-optimizing the AND/OR/XOR neural network and its randomly transformed counterpart on different tasks. Geometrically, this can be described as finding two points in the functional solution spaces of the two networks that are linked by the transformation vector. We move towards extending this approach to more complex datasets by showing task-switching between two sets of 5 MNIST classes and comparing performance to task switching implemented as extra inputs.
Submission Number: 129
Loading