Neural networks: insane in the membrane
summary
Shows hands-on experiments that alter a pretrained image classifier and measures what actually breaks: swapping bright versus dark cat photos and small injected label shifts flips captions and class probabilities, turning near-perfect validation into failure modes as low as around 5% on targeted subsets, while other manipulations push apparent accuracy toward 100% on biased data. Demonstrates dataset-sourced failures like the tank/forest and ImageNet caption distortions, reproduces a chatbot failure from toxic online data, and outlines simple mitigations and diagnostics.
AI-written summary · about
chapters (6)
- 0:00Introduction to neural networks
- 3:40Training neural networks
- 14:22Demo: Image categorization
- 20:06Bias in supervised learning
- 26:23Real-world applications
- 32:49Q&A and closing
related talks
48mKeynote – Sameer Singh
39mMachine learning failures - for art!
25mTricking Neural Networks : Explore Adversarial Attacks
27mUnderstanding Deep Neural Networks
34mDespicable machines: how computers can be assholes
13mHow to trick a neural network!
31mThe dark side of deep learning
40mMachines are Learning