Neural Networks And Spurious Correlations
summary
Neural nets often latch onto easy but unstable cues, producing high worst-group error when spurious attributes (like background color) are easier to learn than foreground features and majority groups vastly outnumber minorities. Re-weighting minority examples usually fails for large-capacity models; removing many majority examples — sub-sampling majority (even ~90% in experiments) — restores large-model performance. A heuristic using early-stop misclassified points to identify and upweight hard examples recovers roughly 80% of the gap on synthetic and CelebA-style benchmarks.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:15Typical machine learning pipeline
- 3:31Taming the neural network beasts
- 4:08In this talk...
- 4:59Models latch onto spurious correlations
- 10:31Improving worst-group error Simple and well-studied approach: reweighting loss based on group proportion
- 11:13Talk outline
- 14:03When do larger models have high worst-group error?
- 18:48Can larger models obtain low worst-group error?
- 20:22Empirical performance of subsampling
- 25:01Summary (spurious correlations)
- 27:42Why do networks pick up on spurious correlations?
- 29:10Simplicity bias of neural networks
- 31:19How to overcome simplicity bias
- 34:10Just train twice - results
- 35:58Parting thoughts
related talks
06mNeural Networks Demystified: Part 7
34mNeural networks: insane in the membrane
48mKeynote – Sameer Singh
01h 08mUnwanted Learning: What It Is, Why It Matters, and What To Do About It
30mMachine Learning Techniques for Class Imbalances & Adversaries
23mEmbracing Noise- How Data Corruption Can Make Models Smarter
32mThe Dangers of Mindless Forecasting
34mDespicable machines: how computers can be assholes