Hello Deep Learning
summary
A 1,500-line 1500-line implementation builds an OCR that trains on MNIST-style handwritten digits without a GPU and actually reads messy handwriting; a trivial average-difference classifier gets 97% on 3 vs 7 and the learned linear-plus-ReLU network reaches roughly 90% accuracy across ten digits but collapses under small shifts unless augmented. The code shows autograd, matrix multiplies, softmax, and Adam-style optimization in practice, and points to compact tools like whisper.cpp and llama.cpp for local models.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:03Deep learning from scratch
- 5:21Hello deep learning project
- 5:52Attention is all you need
- 11:41Weirdly impressive
- 15:34That can't possibly work
- 16:49This really is magic
- 17:11The loss function
- 20:00Layers
- 28:07Auto-gradient
- 28:26Gradient descent with momentum
- 31:13Percentage of digits classified correctly
- 34:51The rules
- 37:12Parting thoughts
- 42:05Useful links
- 43:45Outro
related talks
40mBuilding a Handwriting Recognition System
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask
01h 06mDeep Learning with TensorFlow
01h 07mDeep Learning with TensorFlow
57mWriting a Neural Net from Scratch
51mDeciphering doctor's handwriting with deep learning
55mDeep Learning with CNTK and F#
44mBuilding Handwriting Recognition System for the NY Times Crossword