Figure 6

Investigation of the Impact of Altering the Scale-specificity of the Parameters π′s. Two studies are conducted where the first study (a) looks at the impact of changing the scale-specificity of the π′s and the second study (b) looks at the impact of a larger-scale alteration when the components of individual π matrices are themselves altered. (a) This sub-figure depicts the result of swapping each of the π _i matrices with the last matrix, i.e., with π₁₆. For example, the first data point shows the mean performance when π₁ is swapped with π₁₆, the second data point corresponds to the case when π₂ is swapped with π₁₆ and so on for each of the 16 π _i matrices being swapped with the last matrix π₁₆. Thus the last data point corresponds to the case when the original order was retained. The error bars represent the standard deviation. The results suggest scale-specificity of the learned parameters, i.e., in the sense that the performance degrades drastically if the π matrices of one scale are swapped with a π matrix of a distant scale. (b) The histogram of Pearson correlations depicts the performance when all the π _i matrices are stacked together and the rows of the resulting stacked Π-matrix are swapped randomly 250 times. Such global alteration drastically degrades the performance. Together, these results indicate that the learned parameters do not predict FCs by chance but play a crucial role in the MKL model.