Transfer Learning

Making Deep Learning useful everyday:

• ImageNet has > 1 Million images over 1000 categories
• Labeling data is expensive
• But we do not need to learn from scratch ...

Two predominant ways of doing transfer learning:

• Feature extractor (Network as a feature representation)
  • Case Study: Go from ResNet (trained on ImageNet) to separating SeaBass and Salmon images
• Fine-tuning
  • Reduce learning rate
  • Fix initial layers

How do you decide the strategy?

Source Domain: Type of data on which an existing network is trained

Target Domain: Type of data on which we want to make classification, detection, segmentation

1. Trying to do tumor detection from MRI Scan but you only have 10 images

• Get more data or a network pre-trained on MRI scans

2. Trying to do tumor detection from MRI scan but you can get a lot of images from a hospital

• Train from scratch (if you have sufficient data)
• Fine-tune the network

3. Amazon Go wants to detect their grocery objects and have lot of images

• Fine-tune the network
• Even feature extraction will give reasonable performance

4. Amazon Go wants to add a new grocery object and they can only take 30 images

• Use the existing network as a feature extractor

<< Keras CNN architectures >>