Network Dissection

Introduced by Zhou et al. in Interpreting Deep Visual Representations via Network Dissection

Network Dissection is an interpretability method for CNNs that evaluates the alignment between individual hidden units and a set of visual semantic concepts. By identifying the best alignments, units are given human interpretable labels across a range of objects, parts, scenes, textures, materials, and colors.

The measurement of interpretability proceeds in three steps:

  • Identify a broad set of human-labeled visual concepts.
  • Gather the response of the hidden variables to known concepts.
  • Quantify alignment of hidden variable−concept pairs.
Source: Interpreting Deep Visual Representations via Network Dissection


Paper Code Results Date Stars


Task Papers Share
Decision Making 1 10.00%
Adversarial Defense 1 10.00%
Classification 1 10.00%
General Classification 1 10.00%
Image Classification 1 10.00%
Adversarial Robustness 1 10.00%
Image Generation 1 10.00%
Image Retrieval 1 10.00%
Object Detection 1 10.00%


Component Type
🤖 No Components Found You can add them if they exist; e.g. Mask R-CNN uses RoIAlign