Figure 8. Computing the optimal nonlinearity for each alternative nonlinear classifier axis.

The distributions of the responses to each condition, projected along each axis, were modeled as Gaussian distributions and the responses to the set of target matches and set of distractors were each modeled by mixtures of these Gaussian distributions. To compute an optimized nonlinearity that maximally separated the means of the target matches and distractors, we computed the log likelihood ratio between the mixture of Gaussians for matches and the mixture of Gaussians for distractors. This optimized nonlinearity was computed using the training data, and was applied to the test data following the application of different linear projections to obtain the results in Figure 3c.