Figure 5.

Feature extraction from natural images by the present learning rule. (A) Randomly taken 12× 12-pixel patches from gray-scaled, low-pass filtered and whitened natural images are used as inputs to a recurrent neural network through ON and OFF cells. (B) $\widehat{I}$_gauss is calculated during the course of learning. (C) STAs of input images are displayed as 10× 10 small image patches in a gray-scale (1: black, 255: white), representing the receptive field properties of the neurons in the recurrent network after learning. Each small patch corresponds to an STA due to the firing of a neuron in the recurrent network. (D) The connection weight matrix after learning is displayed on a color scale. The following parameters have been used in this simulation: ϵ = 0.02, c_η = 250.0, c_κ = 150.0, c_ζ = 1000.0, p₀ = 0.0015, p_max = 0.95, N = 100, τ = 5, T = 50000.