Figure 5. Results of ICA in the undercomplete condition.

Thirty-two neurons were used to separate two sources. In each panel, the horizontal and vertical axes respectively represent the first and second elements of two-dimensional vector k_i (i = 1, … , 32), which respectively represents the responsiveness of neuron i to the two sources. ICA is successful if k_i^T ∝ (1, 0) for some i and k_i^T ∝ (0, 1) for others. Initially, vectors k₁, …, k₃₂ are randomly sampled on a unit circle (top left). The EGHR successfully performed ICA as indicated by k₁, …, k₃₂ directed either along the horizontal or vertical axis (top center). On the other hand, the other learning rules did not achieve ICA (other panels). See Methods for other simulation details.