AI Paper Atlas

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

IJCV 2022 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 6 citations (OpenAlex)

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 changedWhereMeasuredEffect
use hierarchical probabilistic pairing (with global concepts) instead of original Network Dissection IoU pairingface-centric ResNet models · interpretable-unit coverage across Block3/Block4 layerscoverage of interpretable units: around 50-60% -> roughly 70-80%Improved
score concepts probabilistically within a facial region rather than reporting only top-IoU conceptGender model Unit 142 · Cheek region conceptsconcept probability: 5 o Clock Shadow 0.1884 vs Rosy Cheeks 0.2276 despite IoU 0.1109 vs 0.1914Improved
increase eyeglasses bias P from 50 to 100 in gender training setResNet-50 gender classifier · Eyeglasses-biased gender classificationinterpretable units for Eyeglasses: more than thrice the number of units as the balanced modelImproved
increase eyeglasses bias P toward 100ResNet-50 gender classifier · Eyeglasses-biased gender classificationinterpretable units in eye regionImproved
increase eyeglasses bias P toward 100ResNet-50 gender classifier · Eyeglasses-biased gender classificationinterpretable units in cheek regionHurt
intermediate bias levels of eyeglasses (between P=50 and P=100)ResNet-50 gender classifier · Eyeglasses-biased gender classificationinterpretable units for EyeglassesNo clear effect
increase grayscale bias P from 50 to 100ResNet-50 gender classifier · Grayscale-biased gender classification, 4 layersnumber of biased units (P>0.55)Improved
compare last layer to earlier layers under color biasResNet-50 gender classifier · Grayscale-biased gender classificationpercentage of biased unitsImproved
train FairFace model on smaller dataset with 60% female skew in smile classSmile classifier · CelebA smileunits paired with a gender: 60% females vs 40% malesHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy