Introduction to Active Learning

When we use some traditional supervised learning methods to classify, it is often the larger the size of training samples, the better the effect of classification. However, in many real-life scenarios, it is difficult to obtain a marker sample, which requires experts in the field to manually label it, and the time and cost of the economy are significant. Moreover, if the training sample is too large, the training time will be more expensive. So is there a way to use fewer training samples to get a better classifier ? Active Learning (Active Learning) provides us with this possibility. Active learning uses certain algorithms to query the most useful unlabeled samples, to be labeled by experts, and then to use the sampled sample training classification model to improve the accuracy of the model.

In the process of human learning, we usually use the existing experience to learn new knowledge, and rely on the knowledge gained to summarize and accumulate experience, experience and knowledge are constantly interacting. Similarly, machine learning simulates the process of human learning, uses existing knowledge to train models to acquire new knowledge, and modifies the model through accumulating information to obtain more accurate and useful new models. Different from passive learning passive receptive knowledge, active learning can selectively acquire knowledge,

The model of active learning is as follows:

A= (c,q,s,l,u),

which C is a group or a classifier, andL is used to train annotated samples. Q is the query function, used to never label the sample pool U Query Information large,S is the supervisor, can be u The correct label is labeled in the sample. Learners start with a small number of initial marker samples L , through a certain query function Q Select one or a batch of the most useful samples, and ask the Supervisor to label, Then use the new knowledge gained to train the classifier and make the next round of queries. Active learning is a cyclic process until a certain stop criterion is reached.

just now, the query function q is used to query one or a batch of the most useful samples. So, What kind of samples are useful? What kind of sample does the query function query? among the various active learning methods, the most common strategies for designing query functions are: uncertainty criteria (uncertainty) and differences Guidelines (diversity).

for uncertainty, we can use the concept of information entropy to understand. We know that entropy is the concept of measuring information and the concept of measuring uncertainty. The greater the entropy, the greater the uncertainty, the richer the amount of information contained. In fact, some of the active learning query functions based on uncertainty are designed using information entropy, such as the Entropy value bagging query (Entropy query-by-bagging). Therefore, the uncertainty strategy is to find ways to identify high-uncertainty samples, because these samples contain a wealth of information, for our training model is useful.

So how does the difference between the opposite sex come to understand? One or a batch of samples is queried before or during each iteration of the query function. We certainly hope that the information provided by the sample being queried is comprehensive and that the information provided by each sample is not redundant, that is, there is some difference between the samples. In the case that each iteration takes a single largest sample of information into the training set, each iteration of the model is re-trained, and the new knowledge to participate in the evaluation of the sample uncertainty can effectively avoid data redundancy. But if you query a batch of samples per iteration, you should try to ensure that the sample is differentiated and avoid data redundancy.

The following is a summary of two articles:

"1" settles B. Active Learning Literature Survey[j]. University of Wisconsinmadison, 2010, 39 (2): 127–131.

"2" Fu Y, Zhu X, Li B. A survey on instance selection for active Learning[j]. Knowledge and information Systems, 2013, 35 (2): 249-283.

Introduction to Active Learning