Columbia Engineering researchers led by Nuttida Rungratsameetaweemana published in PLOS Biology a neural network model revealing how the brain routes top-down signals to early visual cortex. The work identifies a specific circuit: inhibitory neurons suppressing other inhibitory neurons passes task context from higher to sensory areas, overturning the textbook view that early regions merely relay raw input.

The team built a biologically constrained recurrent network with excitatory and inhibitory neurons organized in a sensory-higher hierarchy. When inhibitory-on-inhibitory connections were weakened, task-switching collapsed while other connections left performance intact. Mouse visual cortex recordings confirmed silencing these cells impairs context tracking, matching model predictions exactly.

This disinhibitory motif challenges decades of neuroscience dogma. Rungratsameetaweemana's earlier work with hippocampus patients showed early regions store information redundantly. The finding offers a principle for leaner, more adaptive AI — recurrent networks that achieve cognitive flexibility without transformer-scale training data or energy costs.

The lab is now recording from epilepsy patients with implanted electrodes to test these hypotheses against human neural activity during cognitive tasks, moving from model to direct biological validation.