Learning From Machines
summary
Shows how convolutional nets (Inception) reveal layerwise features by running gradient ascent on Inception to produce DeepDream visuals — iterative scaling plus nudging pixels creates zooming, hallucinatory patterns that reliably expose filters. Explains GANs as a generator and discriminator game that makes realistic but flawed faces (extra limbs, counting errors). Gives measurable biology: retina compresses ~100x so optic nerve is ~10 Mb/s, LSD binds the 5-HT2A receptor for ~220 minutes, grid cells map position, and efference copies + cerebellar Smith‑predictor explain why you can’t tickle yourself.
AI-written summary · about
