Hebbian learning of orientation-selective receptive fields with monotonic fall-off of input correlations

Orientation-selective receptive fields (OSRFs) are shaped by specific feedforward connectivity from thalamus to V1 [Reid 1995]. Previous modeling found that OSRFs will develop from a "Mexican-hat" input correlation pattern between same-type (ON-ON, OFF-OFF) and opposite-type (ON-OFF) correlations [Miller 1994]. However, experimental work suggests a monotonic fall-off of this correlation difference.

Here we study a two-layer network model with Hebbian development of the feedforward connectivity. We find analytically that OSRFs will develop with a monotonic fall-off of the correlation difference and intracortical connectivity, if they are sufficiently long-range and homeostatic competition ensures the summed projection strength of each thalamic neuron is conserved. In this analysis plasticity was driven by averages over activity patterns. Implementing this in a network of excitatory and inhibitory cells with plasticity driven by trial-by-trial activities and recurrent connectivity consistent with experimental data, OSRFs still robustly form. Computing experimentally observed OSRFs properties, we find a quantitative agreement between model and experimental data.

This demonstrates a new, robust mechanism for the development of OSRFs matching several experimentally observed properties.