Learning from Machines
summary
Explains how DeepDream on Inception uses gradient ascent to amplify intermediate feature maps, how convolutional filters mirror retinal edge detectors (100 photoreceptors per ganglion) and the optic nerve tops out around 10 megabits, and how recurrent feedback in cortex contrasts with feedforward nets. Describes GAN-generated fake celebrities and adversarial examples, notes GAN failure modes (extra limbs, bad counting), links 5‑HT2A binding to psychedelic visual rumination, cites cerebellum as a Smith‑predictor for efference copies and grid cells for spatial encoding.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 2:24Hallucinations
- 6:08Neural Networks
- 8:02Retina
- 12:27Visual cortex
- 22:27Convolutional Neural Networks
- 23:49generative adversarial networks
- 26:25fallout cow
- 27:32trippy
- 28:45Neurons
- 30:37Efference
- 32:42The Cerebellar
- 36:52Questions
