CanChen ggchen@mail.ustc.edu.cn
Siamese
- Motivation: For few-show learning problems, we can use another loss function if we do not have enough training examples. This loss function can be similarity-based, which gives a big loss if two inputs belong to the same class and a small loss if they belong to different classes.
- Method: The method is quite simple, trying to maxmize the loss for impostor pairs and minimize the loss for genuine pairs. Then the author did many experiments with half genuine pairs and half impostor pairs.
- Contribution: What I learnt is that we can not count on using genuine pairs only since the network can be trained as a constant function.