Learning from machines
summary
Claims that artificial and biological vision share mechanisms and failure modes: DeepDream uses gradient ascent on Inception layers to amplify mid-layer features and produce hallucinatory images, GANs train a generator and discriminator pair to produce realistic fake faces, and adversarial examples reliably fool classifiers. Neurobiology details back this: LSD binds the 5-HT2A receptor in pyramidal cells to amplify feedback in visual cortex, grid cells encode spatial vectors, and the cerebellum acts like a Smith predictor for motor prediction.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:43Who am I
- 0:54What Ive learned
- 2:05Hallucinations
- 7:00Retinas
- 10:38Visual cortex
- 19:44Optical illusions
- 21:04Fake celebrities
- 24:42Neurocognitive precedent
- 27:42Efference
related talks
44mLearning from Machines
37mLearning From Machines
35mLearning From Machines
42mKeynote: Learning from Machines
12mENHANCE! Upscaling Images with Neural Networks
40mVisual concept learning from few images
01h 15mUnderstanding (a bit about) hallucinations in Generative AI
43mHitchhiker’s Guide to Generative Adversarial Networks (GANs)