The Miseducation of This Machine
summary
Argument: current ML often imitates but does not truly learn — data reproduces human bias, algorithms remix patterns, and only humans supply reflection. Demonstrations use convolutional neural network image filters (bike classifier) and a poisoned image that still registers as bike. Hate‑speech work uses supervised learning on human‑labelled corpora, contrasting cheap crowd labels with a trained code book process; that improves inter‑coder agreement but still fails on sarcasm, quotations and context, so automatic detection is useful but imperfect.
AI-written summary · about
chapters (10)
- 0:00Start
- 6:15I. REPRODUCTION
- 8:50II. REMIXING
- 10:18You're not a robot.
- 14:26CONVOLUTIONAL NEURAL NETWORK
- 17:00III. REFLECTION
- 18:26HOW DO MACHINES IDENTIFY HATE SPEECH?
- 22:14WHERE DO MACHINES FAIL?
- 24:22HOW DO HUMANS IDENTIFY HATE SPEECH?
- 30:21WHERE DO HUMANS FAIL?
related talks
34mDespicable machines: how computers can be assholes
47mEthical Machine Learning
39mSecure Machine Learning: Overcoming Algorithmic Bias, Adversarial ML in the Wild West of AI
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
41mGhost in the Machine: The Unexpected Results of Not Understanding Your Users
38mSocial Implications of Bias in Machine Learning
30mWhat machines can teach us about ourselves?
31mThe dark side of deep learning