stem/AI/Neural Networks/CNN/CNN.md

Before 2010s

• Data hungry
  • Need lots of training data
• Processing power
• Niche
  • No-one cared/knew about CNNs

After

• ImageNet
  • 16m images, 1000 classes
• GPUs
  • General processing GPUs
  • CUDA
• NIPS/ECCV 2012
  • Double digit % gain on ImageNet accuracy

Full Connected

MLP

• Move from convolutional operations towards vector output
• Stochastic drop-out
  • Sub-sample channels and only connect some to MLP layers

As a Descriptor

• Most powerful as a deeply learned feature extractor
• MLP classifier at the end isn't fantastic
  • Use SVM to classify prior to penultimate layer

!

Finetuning

• Observations
  • Most CNNs have similar weights in conv1
  • Most useful CNNs have several conv layers
    • Many weights
    • Lots of training data
  • Training data is hard to get
    • Labelling
• Reuse weights from other network
• Freeze weights in first 3-5 conv layers
  • Learning rate = 0
  • Randomly initialise remaining layers
  • Continue with existing weights

!

Training

• Validation & training Deep Learning#Loss Function
• Early
  • Under-fitting
  • Training not representative
• Later
  • Overfitting
• V.Deep Learning#Loss Function can help adjust learning rate
  • Or indicate when to stop training

!