Keynote: Learning from Machines
summary
Shows concrete visualizations from Inception deep dream gradient ascent and explains convolutional filters, receptive fields, and how gradient ascent produces hallucinatory layers; demonstrates GAN outputs of fake celebrities and their quirks while naming the training loop as GAN generator/discriminator and noting failures at global structure and counting; argues 5-HT2A binding by LSD amplifies cortical feedback (5-HT2A LSD receptor); cites a simulated rat-brain chunk and recommends recurrent, looped architectures over feedforward nets.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:42Neuroscience to machine learning
- 1:50Inception hallucinations
- 6:07Human visual processing
- 11:59Visual cortex feedback loops
- 15:54Psychedelics and pattern recognition
- 19:59Dreams and generative adversarial networks
- 26:41Self-tickling and cerebellar prediction
- 31:11Jokes and discontinuity
- 32:13Self as creation process
- 33:26Q&A
