PaperReading20200404

CanChen ggchen@mail.ustc.edu.cn

---

A few days ago, I wrote a post about a survey on active learning but did not finished reading. Today I am going to give some supplementaries in a different format.

 

Active learning

• Analysis: This paper gives us empirical and theoretical analysis about active learning. In the empirical analysis, the labeled instances may be biased towards the model's distribution.In the theoretical analysis, the author used binary search to illustrate the effectiveness of active learning.
• Problem Setting Variants: There are many variants here and I am
  most interested in active clustering. One idea is to cluster data data based on an expected value of information criterion.
• Relationship with Semi-Supervised Learning: Semi-supervised learning and active learning both want to make good use of unlabeled data. Semi-superivised is prone to exploitation while active learning wants exploration.

 

Triplet

• Motivation: Inspired by the Siamese network dealing with data pairs, this paper introduces Triplet network to deal with triples.
• Method: The author proposes use a parameter-shared network to train three items as an instance, in which two belongs to the same class and two does not belong to the same class. Then a similarity between two instances is calculated and the two similarity scores are sent to binary softmax. At last, the author uses MSE to compute the loss between the output and 0-1 vector.
• Contribution: For me it provides a new view to train a neural network.