The Joy of Designing Deep Neural Networks
summary
Company product for turning receipts to JSON moves from a hand‑baked k‑nearest neighbor heuristic with a six‑person labeling pipeline and 60–70% end‑to‑end accuracy to a deep neural network that hits 97–98% on the same data. Practical experiments compare LSTM, convolution, attention, memory networks and Siamese designs, use Inception/ResNet intuitions, and note massive compute examples like 2000 GPUs for 3 months ~$100,000, which shapes what can be tried.
AI-written summary · about
chapters (31)
- 0:00Introduction
- 0:38Programming background
- 1:34First exposure to AI
- 3:27Business degree
- 3:55Joining Sensible
- 4:48Hand baked heuristic algorithm
- 5:28Building datasets
- 7:47Upgrading to deep learning
- 9:33Early days of deep learning
- 10:21My secret
- 11:20Dont know the equations
- 12:24The sigmoid
- 13:08When does it matter
- 13:52What do layers do
- 14:12ALICE TM
- 15:22Convolution
- 15:33Understanding convolution
- 16:23Understanding layers
- 17:34What are layers
- 18:28What are attention layers
- 20:47Architecture
- 21:32Edges and nodes
- 22:47Translation
- 24:49Intuition
- 26:18Gossip
- 27:56Experimentation
- 29:21Neural Network Learning
- 30:42Different Approaches
- 32:31Dark Heart
- 34:18Backpropagation
- 36:42Conclusion
related talks
27mDemystifying Deep Neural Networks
59mDeep Neural Networks for Video Applications
58mDeep Learning in Real Life
54mIntroduction to Understanding Deep Learning
28mDemystifying Deep Neural Nets
42mDeep Learning for Computer Vision
51mGetting Started with Deep Learning for Software Developers
37mWhat's new in Deep Learning?