Learning understandable classifier models.

Receptive fields optimization in deep learning for enhanced interpretability, diversity, and resource efficiency.

“…Therefore, most of the existing DL models can only be used as black-boxes despite their good performance because their knowledge is hidden and can hardly be used to explain their decision making process. Thus limiting their applicability in domains where both justifications of decisions and interpretable inference are required from machines as in medical applications and business intelligence [20].…”

“…Initial number of filters for each layer is shown in the legend [3]. Figure 37 shows the sensitivity of ResNet-56 layers to pruning and it can be observed that layers such as Conv 10,14,16,18,20,34,36,38,52 and 54 are more sensitive to filter pruning than other convolutional layers. Likewise for ResNet-110, the layer sensitivity to pruning is depicted in Figure 38 and it can be observed that Conv 1, 2, 38, 78, and 108 are sensitive to pruning.…”