A role for cortical interneurons as adversarial discriminators
Fig 2
A simple case illustrating the problem of predicting activity in a population with internal connections.
a) We created a toy task in which the goal is to adjust the connections away from neuron A to better model the activity of neurons B and C. The “target population” [B, C] is also driven by an external teaching signal from neuron T. All connections evoke Gaussian-distributed postsynaptic responses, and all neurons sum their inputs linearly. b) When learning the A → B and A → C connections by predicting the postsynaptic neuron (e.g. maximizing log p(B|A), ignoring recurrence), the predictions fail to match the joint distribution. The left plot shows the true joint distribution p(B, C), which is correlated and non-Gaussian due to the B → C connection internal to the target population, while the right plot shows . c) The adversarial strategy successfully aligns the distribution. The discriminator sees B and C and gates plasticity at the predictive connections from A. d) The alignment can be quantified with the KL divergence of the binned histogram between the learned and target distribution. We plot the trajectories from 5 random runs; all networks share the same initialization. e) Another measure of success is the distance of the parameters of the outward connections from A from their optimal value, which we call the parameter error.