Figure S7: Comparison among different statistical classifiers and selectivity criteria

A, B. Comparison of classification performance levels obtained using different statistical classifiers. Throughout the text, we report the performance of a Support Vector Machine (SVM) classifier with a linear kernel. We compared the performance of this classifier against other possible machine learning classifiers (Bishop, 1995; Hung et al., 2005a; Vapnik, 1995). Here we show a direct comparison between the linear SVM (x-axis) and an SVM classifier with a Gaussian kernel (red) or a nearest neighbor classifier (blue). The diagonal dashed line would correspond to identical performance across classifiers. We compare the values only for those electrodes and conditions that yielded a performance above 3 standard deviations of the null hypothesis, which was our threshold throughout the text for statistical significance (see Figure S5). Overall, the Gaussian SVM showed a slightly better performance and the nearest neighbor classifier showed a slightly worse performance than the linear SVM; yet, none of the conclusions in the current manuscript would be significantly modified if we used these other classifiers. In A. we compare the results for multiple IFP response definitions (see Figure S8) whereas in B. we restrict the comparison to the IFP range in the 50 to 300 ms window. C, D. Comparison between selectivity defined by using a linear SVM classifier (y-axis) and using a one-way ANOVA on the IFP responses to define selectivity (Thorpe et al., 1996). The x-axis shows the p value of obtained in the ANOVA analysis in log scale for all the electrodes (C) or only those electrodes that showed p<0.01 (D). The green line shows a linear fit to the data.