Compared with semi-supervision

[czy341181]

I would like to ask what is the difference between active learning and semi-supervision.

I am doing semi-supervised, I read the paper and the code and didn't find difference,
Is active learning and semi-supervised the same thing？ Or What is the main difference between the two, and can I directly use active learning for semi-supervision？

[yuantn]

I would like to ask what is the difference between active learning and semi-supervision.

I am doing semi-supervised, I read the paper and the code and didn't find difference,
Is active learning and semi-supervised the same thing？ Or What is the main difference between the two, and can I directly use active learning for semi-supervision？

The process of active learning can be found in the first paragraph of Section 3.1. The core of active learning is that we first train a model with small amount of data, and then calculate the uncertainty (or other designed metrics) to select the informative samples for the next active learning cycle.

The advantage of active learning over supervised learning is that it can gradually select the samples that it wants to learn from. Conventionally, active learning focuses on how to select the informative samples in each active learning cycle, and tries to reach the upper bound of performance (i.e., the performance of fully supervised learning) as soon as possible.

However, in my opinion, the semi-supervised learning solve the problem in another way. It tries to mine and utilize unlabeled samples in a static perspective but not a dynamic perspective.

I think that our work MI-AOD cleverly combine the active learning with semi-supervised learning. That is, we use semi-supervised learning (or its key idea) to learn with limited labeled data and enough unlabeled data, and use active learning to select informative unlabeled data and annotate them. This is the trend of the recent research in active learning, and use active learning for semi-supervised learning is also a good idea.

[czy341181]

So after active learning to select those informative samples, you still need to humanly annotate it before next cycle?

[yuantn]

So after active learning to select those informative samples, you still need to humanly annotate it before next cycle?

In actual industrial applications, the human labeling after each cycle is required. But in the experiment, we only need to enable the label and annotation of the "unlabeled" samples, because the annotations of all samples in the datasets are available.

[czy341181]

got it. Thanks for your patient reply!