CanChen ggchen@mail.ustc.edu.cn
Recently I have been working on active learning and today I am going to share my reading paper about an survery on active learning. I have not finished reading this paper.
Active Learning Literature Survey
- Motivation: During training we do not need all the data since labels are difficult to get in some cases. Some specific instances and labels, which contain enough information and can represent lots of other data points，with which we can train a good model.
- Method: Active learning includes three scenarios: membership query synthesis, stream-based sampling and pool-based sampling. Membership query synthesis query data points by generating data with the help of learnt function. For the stream-based sampling and pool-based sampling, the difference is that the previous one decides to accept the instance only considering this point while the latter one sample data points with the knowledge of other data points in the pool. For query frameworks, there are uncertainty sampling, query-by-committe, expected model change, expected error reduction, variance reduction and density weighted methods. Except for the last one which try to find the representative instances, others try to find informative instances.
- Contribution: Active learning is a commmon learning algorithm and the review gives me a big picture of active learning.