Deep Learning · TEE · chapter

CNN architectures and transfer learning

AlexNet, Inception, ResNet, bottlenecks and transfer learning. Study their structure, design purpose and calculations.

1. AlexNet and the reason for its components
2. Inception examines several spatial scales
3. Why one by one convolution saves work
4. Drill 39 Bottleneck cost with new values
5. ResNet adds a shortcut
6. Drill 40 Residual addition
7. Drill 52 Projection shortcut
8. Transfer learning from a small dataset
9. Inception concatenation and residual addition
10. Drill 51 Inception concatenation and parameter count
11. Applications of convolutional neural networks
12. Reading LeNet and AlexNet layer diagrams
13. Top-1 and top-5 error: what an ImageNet result means
14. VGG16: small filters in a regular sequence
15. Global average pooling: one average for each channel
16. The faculty's 480-channel bottleneck numerical
17. Batch normalization inside a residual network
18. Residual-block order and architecture qualifications
19. Choosing frozen layers from dataset size and similarity
20. Four ways to reuse a pretrained model
21. F25 Q2b
22. F25 Q3b
23. R26 Q2b
24. R26 Q7b
25. F23 Q4a
26. F23 Q5a
27. F23 Q6b
28. R24 Q6b
29. F22 Q4a
30. R23 Q4b