(A) Top: Abstract characterization of a communication channel. A stimulus is sampled from an information source and passed through a noisy communication channel , which outputs a stimulus reconstruction . The reconstruction error is quantified by a distortion function, . Bottom: Circuit architecture implementing the communication channel. Input neurons encoding the negative distortion function provide the driving input to output neurons with excitatory input ui and global feedback inhibition . Each circuit codes a single stimulus at a fixed retinotopic location. When multiple stimuli are presented, the circuits operate in parallel, interacting only through a common gain parameter, . (B) Tuning curves of input neurons encoding the negative cosine distortion function over a circular stimulus space. (C) Rate-distortion curves for two different set sizes ( and ). The optimal gain parameter is shown for each curve, corresponding to the point at which each curve intersects the channel capacity (horizontal dashed line). Expected distortion decreases with the information rate of the channel, but the channel capacity imposes a lower bound on expected distortion. (D) Example spike counts for output neurons in response to a stimulus (, vertical line). The output neurons are color coded by their corresponding input neuron (arranged horizontally by their preferred stimulus, for neuron ; full tuning curves are shown in panel B). When only a single stimulus is presented (), the gain is high and the output neurons report the true stimulus with high precision. (E) When multiple stimuli are presented , the gain is lower and the output has reduced precision (i.e., sometimes the wrong output neuron fires).