Compositional computations in the primate frontal cortex

Frontal cortex activity in trained monkeys separates compositional generalization into spatial and temporal components, with novel input combinations producing accurate responses only when shared timing dynamics are preserved.

• Why it matters: It remains unclear how learned neural computations are carried to untrained states without relearning, and why generalization can fail in predictable ways.
• What they did: Monkeys learned to combine a start location with optic-flow direction and duration, while neural recordings and circuit modeling tested how input representations, recurrent dynamics, and readout jointly implement the computation.
• The result: Novel combinations that maintained the learned organization elicited accurate responses, but departures predicted systematic behavioral biases that grew with distance from trained conditions, mirrored in humans under an unfamiliar sensory-spatial mapping.

AI summary of the abstract · may contain errors · not medical advice · read the paper.