Learning the Similarity for Classification

What if your dataset is small and/or doesn’t cover the full variability of the world ? Traditionally, in this case one has to:

• Regularize harder;
• Use simpler model;
• Perform data augmentation;
• Whatever;

to reduce overfitting.

The idea is to:

• leverage existing large scale image database for pertaining;
• Solve theoretically more difficult problem: learn similarity function in the space of arbitrary yet real images;
• Fine-tuning for the target dataset;
  
• Compute stochastic similarity measure with each class in test time.