Pattern Analysis: Learned vs. Engineered
summary
Deep convolutional nets like AlexNet cut ImageNet error from ~25% to ~15% in 2012 and learn interpretable first‑layer kernels that mirror Gabor filters instead of handcrafted color averages. Visualizations show neurons that fire on faces, text, legs and wheels; some learned filters are flat and useless until models are tuned. On nonvisual data a recurrent model trained on MIT Reality Mining maps weekdays vs weekends into vector distances, revealing useful learned representations.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:21What is deep learning
- 3:05Zero networks
- 7:39Deep representation learning
- 9:18Main points
- 10:31Filters
- 18:16Location Data
