Fig. 4. Linear classifiers and the NNC did not require the use of all available glomeruli to reach plateau performance.

A Example of regressor weights for the linear SVM as the sparseness constraint is changed. B The SVM regressor weights were applied to test set mixtures where methyl pyruvate was the novel background odor. C and D Logistic regressor weights were also calculated from the training set and were applied to the test set. E Number of glomeruli used in the NNC was changed by thresholding based on the auROC of the training set of individual glomeruli. Similarity matrices changed as the threshold was changed. The red squares indicate the best match to the training set. F Performances of the linear classifiers and the NNC as a function of the number of ROI included calculated for 32 recording sessions, 6 WT mice. Vertical error bars are the s.e.m. of the performance of the classifiers and horizontal error bars are the s.e.m. for the number of glomeruli used for the classifiers.