The dark side of deep learning
summary
Five live demos show concrete failures: word2vec / Gensim reveals gendered associations that map “woman” to homemaker roles; adversarial examples generated with Foolbox create universal adversarial noise that reliably misclassify images and audio; membership inference and model inversion recover training data and can reconstruct photos from federated updates; ultrasonic “dolphin” commands activate speech systems without human hearing; and operational hidden technical debt from entangled data pipelines breaks systems months later.
AI-written summary · about
chapters (7)
- 0:00Introduction and background
- 3:56Deep learning is not magic
- 6:25Discrimination in NLP models
- 15:33Real-world model vulnerabilities
- 21:58Data privacy and security risks
- 27:35Technical debt in deep learning
- 30:02Q&A and closing
