1. When you are migrating to learn?
Most machine learning algorithms are the features of the training data and test data, assuming that the distribution of the same. However, this was often not feasible in the real world. For example we have a task we want to classify, but this task is not sufficient data (in the migration study also referred to the target domain) , but there are a large number of relevant training data (in the migration study also referred to as the source domain ) , but the different characteristics of test data classification tasks required to conduct this training data and the distribution (eg speech emotion recognition, voice data of a language adequate, however, required to classify the task of emotion but extreme lack of data), in in this case, if appropriate transfer learning methods can be used, you can greatly improve the classification results of the sample is not sufficient task. That is, it usually refers to the ability to migrate to the new knowledge environment, which is often referred to as transfer learning. Sometimes if the migration will result in a negative migration inappropriate, for example, may lead to negative transfer when the irrelevant task of the source and target domains.
2. Migration learn what migration?
Some features are unique to the individual in some learning tasks, these features can not be migrated. And some have characterized the contribution of all individuals, these can be migrated. ( A brief summary is not unique to the individual characteristics of the sample migration, the migration characteristics of all individuals who have contributed. )
3. Migration kind of learning
Inductive transfer learning ( Inductive Learning Transfer), Direct Push transfer learning (Transductive Transfer Learning), unsupervised transfer learning (unsupervised transfer learning)
Inductive transfer learning ( Inductive Transfer Learning) objectives, tasks and tasks of different original target and source domains may be the same or may be different. which is. This migration has been divided into four learning form 1) instance of knowledge transfer learning (Transferring knowledge of instances) based on the basic idea of transfer learning examples that, although the target portion of the tagged data and source domain training data or training or less is somewhat different, but part of the training data in the target domain should still be present as part of a more suitable training for effective classification model, and adapt the test data. Thus, the goal is to target some with label field training data for test data to identify those instances, and migrate those instances to study the source to the training data. (Do not know if understanding is correct, Originally although the source domain data can not be reused directly, there are certain parts of the data that can still be reused together with a few labeled data in the target domain.) This method is suitable for use with the source data and target data with a tag data portion very close to the case. 2) characteristics of knowledge transfer (Transferring knowledge of feature representations) based transfer learning characteristics of the main features is to find a good representation of different domains minimized. And it can be divided into the adequacy of supervised and unsupervised learning method tagged sample source domain. 3) knowledge transfer parameters (Transferring knowledge of parameters)
Most of migration based learning model parameters are based on different domains shared some parameters or some contribution to the premise of the prior distribution.
4) knowledge of migration (Transferring relational knowledge)
data transfer learning this source and target domains are related. E.g. newworked data and social network data. There is a correlation between their data.
Direct Push transfer learning ( transductive Transfer Learning) Direct Push transfer learning tasks and objectives of the original task is the same, target and source domains may be the same or may be different. It may be divided into different a) Spatial and b) wherein the same space, but a different marginal probability distributions (and domain adaptation like this case) transductive still comprise the instance of knowledge transfer learning transfer learning ( Transferring Knowledge of instances) and characteristics knowledge transfer (Transferring knowledge of feature representations) (this situation is carried out in unsupervised learning model.)
Migration unsupervised learning ( Unsupervised Learning Transfer) unsupervised transfer learning tasks, objectives and tasks of the former type are not identical, and the target domain data as well as data source fields are no labels. For example, recently proposed self-taught clustering, transferred discriminative analysisbelongs to the unsupervised transfer learning. These features and unsupervised learning feature exists only knowledge transfer ( Transferring Knowledge of the Feature Representations).
4. Active Learning Introduction
We use some of the traditional ways of doing supervised learning classification, training is often the larger the sample, the better classification. But in many scenes in real life, it is more difficult to obtain samples of the mark, which requires experts in the field to manual tagging, time spent and economic costs are significant. Moreover, if the training sample size is too large, the time spent training will be more. Is there a way to be able to use fewer training samples to get a better performance of the classifier it? Active Learning (Active Learning) provides this possible for us. Active learning through a certain algorithm queries most useful unlabeled samples, labeled and handed over to the experts, and then use a sample query to train the classification model to improve the accuracy of the model.
In the human learning process, usually using the existing experience to learn new knowledge, but also rely on the knowledge gained to sum up and accumulate experience, continuous interaction experience and knowledge. Similarly, the process of human learning simulation machine learning, the use of existing knowledge to train the model to acquire new knowledge, and through the accumulation of information to correct model to obtain more accurate useful new model. Passive learning is different from the passive acceptance of knowledge, active learning can be selectively acquire knowledge.
The model of active learning
A=(C,Q,S,L,U)，
C is one or a set of classifiers
L is used to train the labeled samples
U is unlabeled sample cell
Q is a query function for labeling the sample reservoir never U selected data
S is a supervisor who may be U marked correct sample Tags
Learner through a small initial sample labeled L to start learning, through a simple query function Q to select one or a group of the most useful samples, and ask those who label supervisor, and then use the new knowledge gained to train the classifier and the next round Inquire. Active learning is a process of a cycle, until a stopping criterion so far.
6. How to select the sample
Active learning is a very important question is how to select the sample. The acquisition unlabeled data samples way type is not the same , the main movable learning learning calculation method may be in minutes as based on stream ( Stream-based), and based on the pool (pool-based) learning strategies, stream-based approach is to turn the incoming each sample query, determines whether it needs to be submitted to the label supervisors; pool is based on a method for caching new sample, when the cache to a certain number, in the selected sample buffer according to standard labeling required supervisors example.
Based pool learning strategy is currently the most widely used, the most adequate theory of strategy.
Active learning cell-based method of selecting a different method of labeling the sample can be divided into a method (based on uncertainty Uncertainty Sampling), vote selection method (Query-By-Committee), a desired change in the model (Expected Model Change), error reduction desired (expected Error reduction), the principle of reducing the variance (variance reduction) and the weight density (density-Weighted method, ), and other methods . One of the most commonly used method is based on uncertainty, active learning is the most uncertain select those samples submitted to the supervisors who are labeled. This method is very convenient option strategy is a probability model.