Fig. 8: Stimulus sampling by a network is reflected in the internally generated differential correlations, whose impact differs from differential correlations inherited from feedforward inputs.
From: Sampling-based Bayesian inference in recurrent circuits of stochastic spiking neurons
a Stimulus sampling via spike generation causes the population firing rate to fluctuate along the stimulus subspace (x-axis). b The pattern of internally generated differential correlation in a network implementing sampling composed of neurons with Gaussian tuning. c Internally generated differential correlations in such a network increase with recurrent weight, wE. d The rate in feedforward input decreases the externally generated correlations, and increases the mutual information between the feedforward inputs and latent stimulus. e Recurrent network weights increase internally generated differential correlations. Mutual information between stimulus and feedforward inputs changes non-monotonically with recurrent weight. The direction of arrows indicates the predicted direction of change of the recurrent weights after an animal is retrained using a new stimulus set with different correlations compared to the reference stimulus set.
