Code First, Math Later: Learning Neural Nets Through Implementation and Examples
summary
Uses Keras code to build actual models rather than theory: a convolutional net trained on 50,000 images with conv + pooling → flatten → dense → softmax, and an LSTM embedding + recurrent stack for tweet relevancy with a 10,000-word vocabulary and 32-token padding. Shows concrete debugging: per-image probability outputs reveal confusion, demonstrates dimension-mismatch fix (return_sequences=True), and inspects learned embeddings to confirm grammatical clusters — models run and reveal practical limits.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 2:06Why learn neural networks through implementation
- 3:10Deep learning fundamentals
- 6:17Convolutional networks for images
- 8:44Recurrent networks for sequences
- 10:50Keras API
- 13:41Implementing object detection CNN
- 25:07Document classification with LSTM
- 31:30Debugging architectures
- 34:47Visualizing word embeddings
- 36:54Conclusion and key takeaways
related talks
44mMachine Learning Exposed
51mMachine Learning Exposed Deep and Reinforcement Learning
54mNeural networks by example
35mDeconstructing Deep Learning
36mNeural Networks
01h 29mHands-on introduction to Deep Learning with Keras and Tensorflow
32mA word is worth a thousand pictures: Convolutional methods for text
55mA Tour of Deep Learning With C++