Active Learning-1

Http://videolectures.net/icml09_dasgupta_langford_actl/

??

Objective:

Tag as few as possible to obtain the best training results.

In general, we hope that the model we train can provide the best performance for natural distribution, that is, truly underlying distribution.

??

This can be done for the random extraction and labeling of data sets. However, for active learning, such as the above strategy, we tend to select points close to the classification surface, then the data points in our training set will be concentrated on both sides of the classification surface. That is to say, as we continue to iterate our active learning strategy, the training set we select will be farther and farther away from truly underlying distribution.

??

Do we really need to train optimize for such a training set?

??

This leads to the biased sampling

??

The labeled points are not representative of the underlying distribution

??

??

Q: What is the essence of active learning? Active Learning is to select choosing instance instead of randomly selecting instance

A: We want to select the most informative point. At the same time, we hope that the selected training set will play the best in the random data, that is, truly underling distribution, therefore, it seems that there are some tradeoff or conflict, so we want to solve this problem here.

??

??

Here, the active learning strategy will mistake 5% green as red .. But this is not the case.

??

Therefore, the Active Learning Algorithm here is not consistant.

??

Definition of consistant:

When the point set you selected approaches an infinite number, the model trained from the corresponding training set should approach the optimal one.

??

Even with infinitely extends labels, converges to a classifier with 5%

Error instead of the best achievable, 2.5%. Not consistent!

??

??

Can adaptive query solve this problem?

??

There are two distinct narratives for explaining how adaptive

Querying can help.

Case I: Exploiting (cluster) Structure in Data

Case II: efficient search through hypothesis Space

??

Case I: Exploiting (cluster) Structure in Data

??

The problem here is that the definition is unclear, and the clustering itself has a granularity problem. Five clusters, six clusters, and seven clusters ...?

In addition, the labels in each cluster may be inconsistent.

??

Active Learning-1