Post-hoc Self-explanation of CNNs

Dimensions overall score 4.0

GitHub Code Pulse

No public code linked for this paper yet.

• Standard CNNs can be mathematically reinterpreted as Self-Explainable Models (SEMs), but their built-in prototypes do not accurately represent the data on their own.technical 90%
  Evidencepartial

  Although standard Convolutional Neural Networks (CNNs) can be mathematically reinterpreted as Self-Explainable Models (SEMs), their built-in prototypes do not on their own accurately represent the data.

  Implicationpartial

  Directly stated in the abstract and repeated in the analysis section.

  Verificationpartial

  partial

• Replacing the final linear layer of a CNN with a k-means-based classifier addresses the prototype representation limitation without compromising performance.method 85%
  Evidencepartial

  Replacing the final linear layer with a k-means-based classifier addresses this limitation without compromising performance.

  Implicationpartial

  Directly stated in the abstract and supported by performance data in Table 2.

  Verificationpartial

  partial

• Using shallower, less compressed feature activations (e.g., from blocks B234) for explanations results in a trade-off between semantic fidelity and a slight reduction in predictive performance.result 80%
  Evidencepartial

  Empirical evaluation with a ResNet34 shows that using shallower, less compressed feature activations, such as those from the last three blocks (B234), results in a trade-off between semantic fidelity and a slight reduction in predictive performance.

  Implicationpartial

  Explicitly stated in the abstract and analysis, with performance data in Table 2 showing reduced accuracy for B234.

  Verificationpartial

  partial

• The k-means-based explanation method can generate concept-based explanation maps by leveraging the spatial consistency of convolutional receptive fields.method 75%
  Evidencepartial

  The latter approach leverages the spatial consistency of convolutional receptive fields to generate concept-based explanation maps, which are supported by gradient-free feature attribution maps.

  Implicationpartial

  Strongly implied in the analysis section describing the approach, though not a direct quote.

  Verificationpartial

  partial

• The proposed method (Ours B4) achieves 100.0% train accuracy and 75.2% test accuracy on a specific dataset, matching the performance of the backbone network.result 90%
  Evidencepartial

  Ours B4 100.0 99.4 94.0 85.5 100.0 75.2

  Implicationpartial

  Direct numeric evidence from Table 2 in the analysis section.

  Verificationpartial

  partial

• Network Dissection for the final layer of a ResNet34 (B4) involves comparing 512 neurons, which are often redundant according to the authors.limitation 85%
  Evidencepartial

  For the final layer of a ResNet34 (B4), this involves comparing 512 neurons, which, as reported by the authors, are often redundant.

  Implicationpartial

  Direct quote from the analysis section referencing prior work.

  Verificationpartial

  partial

• ACE (Automated Concept Explanation) explains image regions rather than internal concepts because its SLIC segmentation operates independently of the encoder's learned representations.limitation 80%
  Evidencepartial

  Since SLIC segmentation operates independently of the encoder’s learned representations, it explains image regions rather than internal concepts.

  Implicationpartial

  Direct quote from the analysis section critiquing prior work.

  Verificationpartial

  partial

Startup potential card