Assess interpretability

[gcastro-98]

Enhance the decoder CNN by explaining its decisions by:

• Layer Wise Relevance Propagation: there are several pytorch implementations. For instance:
  • Tutorial to apply them to VGG-16 and other arquitectures
  • A different flavour also appliable tto pre-trained large VGG models of torchvision
  • Other different suggestions: note residual connections (e.g. those of ResNet) may give continuity problems
• Gradient-based localization (Grad-CAM):
  • There are plenty of guides out there but OmniXAI provides some interesting implementations
• Other explainability flavours different than pixel-attribution methods:
  • Concept Activation Vectors (CAV): check, for instance, captum implementation
  • Counterfactual explanations
  • Contrastive explanations
  • LIME for image classification