Randomly connected networks generate emergent selectivity and predict decoding properties of large populations of neurons
Fig 1
Overview: Simulation of an input frequency discrimination task in a recurrent network of excitatory and inhibitory neurons with random connectivity.
A: Recurrent E-I network of randomly connected neurons. Synaptic weights are heterogeneous (line thickness), and connectivity in the network is random. Inputs to the network have the same pattern of spatial activation but different temporal features. Readouts of network activity (from all neurons, or restricted to excitatory or inhibitory neurons) discriminate between stimulus categories. We then analyze the distributions of readout weights. B: Distributions of non-zero inhibitory (shaded red) and excitatory (shaded blue) synaptic weights. 20% of the weights are non-zero. C: Inputs to the network are pulses arriving at random times with low (top row, yellow) or high (bottom row, red) average frequency. D: Histogram of population average network activity over trials for low- and high-frequency stimuli. Average network activity is not informative of input categories.