Research map / Label-efficient and robust learning / Recognition and long-tail
Interpreting Face Inference Models Using Hierarchical Network Dissection
Divyang Teotia, Àgata Lapedriza, Sarah Ostadabbas
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hierarchical probabilistic pairing (with global concepts) instead of original Network Dissection IoU pairing | face-centric ResNet models · interpretable-unit coverage across Block3/Block4 layers | coverage of interpretable units: around 50-60% -> roughly 70-80% | Improved |
| score concepts probabilistically within a facial region rather than reporting only top-IoU concept | Gender model Unit 142 · Cheek region concepts | concept probability: 5 o Clock Shadow 0.1884 vs Rosy Cheeks 0.2276 despite IoU 0.1109 vs 0.1914 | Improved |
| increase eyeglasses bias P from 50 to 100 in gender training set | ResNet-50 gender classifier · Eyeglasses-biased gender classification | interpretable units for Eyeglasses: more than thrice the number of units as the balanced model | Improved |
| increase eyeglasses bias P toward 100 | ResNet-50 gender classifier · Eyeglasses-biased gender classification | interpretable units in eye region | Improved |
| increase eyeglasses bias P toward 100 | ResNet-50 gender classifier · Eyeglasses-biased gender classification | interpretable units in cheek region | Hurt |
| intermediate bias levels of eyeglasses (between P=50 and P=100) | ResNet-50 gender classifier · Eyeglasses-biased gender classification | interpretable units for Eyeglasses | No clear effect |
| increase grayscale bias P from 50 to 100 | ResNet-50 gender classifier · Grayscale-biased gender classification, 4 layers | number of biased units (P>0.55) | Improved |
| compare last layer to earlier layers under color bias | ResNet-50 gender classifier · Grayscale-biased gender classification | percentage of biased units | Improved |
| train FairFace model on smaller dataset with 60% female skew in smile class | Smile classifier · CelebA smile | units paired with a gender: 60% females vs 40% males | Hurt |
Closest papers by meaning
- UniF$^2$ace: A $\underline{Uni}$fied $\underline{F}$ine-grained $\underline{Face}$ Understanding and Generation Model (ICLR 2026 · similarity 0.86)
- Network Dissection: Quantifying Interpretability of Deep Visual Representations (CVPR 2017 · similarity 0.86)
- FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs (CVPR 2025 · similarity 0.85)
- A Convolutional Neural Network Interpretable Framework for Human Ventral Visual Pathway Representation (AAAI 2024 · similarity 0.84)
- GroundingFace: Fine-grained Face Understanding via Pixel Grounding Multimodal Large Language Model (CVPR 2025 · similarity 0.84)
- SegFace: Face Segmentation of Long-Tail Classes (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
