Learning Orthogonalizes Visual Cortical Population Codes


September 22, 2020 - 1:00pm - 2:00pm
About the Speaker
Kenneth Harris
Speaker Title: 
Speaker Affiliation: 
University College London

The response of a neuronal population to a sensory stimulus can be summarized by a “rate vector” in a high-dimensional space. Learning theory suggests that the brain should be better able to produce distinct behavioral responses to two stimuli when the rate vectors they evoke are close to orthogonal.


To investigate how learning modifies population codes, we measured the orientation tuning of 5,000-neuron populations in visual cortex before and after training mice on a visual orientation discrimination task. Surprisingly, training reduced the number of neurons preferring task stimuli, most strongly amongst weakly-tuned neurons. These multifarious changes in single-cell tuning reflected a simple change at the population level: sparsening of population responses to the task stimuli, resulting in orthogonalization of their rate vectors.


To understand potential mechanisms of this effect, we built a computational model. This model required no synaptic plasticity: instead, sparsening and orthogonalization of the codes was produced simply by modulation of neuronal F-I curves. Nevertheless, the model accurately predicted how tuning curve plasticity for all groups of neurons.


We conclude that training orthogonalizes the population codes for task-relevant stimuli, and that this may be achieved by simple network mechanisms such as changes in inhibition, without requiring synaptic plasticity coordinated at a network level.