📗 -> Lecture Date: Name
🎤 Vocab
Batch Normalization (BN)
Residual Net (ResNets)
❗ Unit and Larger Context
Small summary
✒️ -> Scratch Notes
Batch Normalization
Two modes of BN:
- Train mode
are functions of x; backprop gradients
- Test mode
are pre-computed on training set
Deep Residual Learning
Deep ResNets can be trained without difficulties
They don’t show the lower training error and test error that normal deep nets do
Packages:
2 main deep learning packages
- TensorFlow
- PyTorch
Breaking CNNs
Humans are susceptible to visual illusions, and CNNs are no different
We might not be able to detect this, as what they can detect is different. Noise is particularly effective.
- Paper
- Github: Breaking ConvNets
Do gradient ascent, to maximize loss instead of minimize
Defense
Train the net on noisy images, this is called adversarial training
🔗 -> Links
Resources
- Put useful links here