Machine learning failures - for art!
summary
A set of concrete failures lets you pick: a keycard experiment where a cookie robot gets let into dorms 19% then 76% of the time; recurrent nets that invent color and ice-cream names and transfer metal-band noise into flavors; image classifiers that hallucinate sheep on grassy fields and a name classifier with a princess effect (100% Pony); a GAN trained on 10,000 tulip photos that succeeds when narrow and fails when broadened; Skynet knitting that outputs tentacles and required human debugging.
AI-written summary · about
chapters (41)
- 0:00Introduction
- 0:34Example of an experiment
- 2:27Example of charming humans
- 3:53Limited information
- 4:10Experiment
- 6:18Cookies
- 6:50Surface appearances
- 7:12Puns
- 7:49Neural net jokes
- 8:36April Fools jokes
- 9:21Limited memories
- 9:38Long memories
- 10:44Unfair situation
- 11:23Pony classification
- 12:57Application image recognition
- 17:22Fun to enjoy
- 17:42Problem is too broad
- 19:06Another example
- 20:36For example
- 20:54Several in fact
- 21:04Consistent
- 21:49Clocks
- 22:05Sheep
- 22:25Wool
- 22:50surrealism
- 23:07sheep surrealism
- 23:18spooky clocks
- 23:32training data
- 25:14boring pictures
- 26:17pizza girl
- 27:11image recognition algorithms
- 28:07adversarial attacks
- 31:23Skynet
- 32:10Knitting example
- 32:43Challenges
- 33:39Knitting
- 35:03Knitting without debugging
- 35:26Tiny baby whales Soto
- 35:53The knitters liked it
- 36:56Perfection is always a bad thing
- 38:08Take home message
related talks
34mNeural networks: insane in the membrane
48mKeynote – Sameer Singh
45mApplied machine learning: a few lessons I learned the hard way
31mThe dark side of deep learning
26mHow Machine Learning Can Help
36mLearning about machine learning from kids
34mDespicable machines: how computers can be assholes
01h 04mCan you trust your (large language) model?