Lightning Talk: Leveraging PyTorch 2.0 for Bias Reduction in AI
summary
Lightning summary: shows concrete steps using PyTorch 2.0 features and third‑party tools to detect and reduce bias — cites the 2018 Gender Shades result (error rates up to 34% higher for darker‑skinned females). Recommends specific implementation moves: use DataLoader.shuffle to avoid order bias, apply torchvision.transforms for augmentation, use regularization (Dropout, BatchNorm) and inspect networks via autograd and layer probes, and run audits with AI Fairness 360 and Captum to measure and explain outcomes.
AI-written summary · about
chapters (9)
- 0:00PyTorch and AI bias
- 1:45AI bias types and examples
- 3:54Bias in facial recognition
- 4:37PyTorch bias tools
- 6:01PyTorch bias features
- 7:25Overfitting and bias reduction
- 8:06Bias detection libraries
- 9:10Balanced datasets and validation
- 9:52Q&A
related talks
31mFlagging *and* Fixing Bias in ML
47mFight AI Bias With… Bias
23mRemoving Unfair Bias in the Machine Learning Pipeline
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
40mEthical AI: Addressing Bias & Algorithmic Fairness
01h 31mHow To Measure And Mitigate Unfair Bias in Machine Learning Models
40mResponsible AI Practices: Fairness in ML
51mDigital Discrimination: Cognitive Bias in Machine Learning