0
$\begingroup$

If I'm generating pseudo-labels that I'm confident are correct for my dataset due to high confidence scores or something else, how can I expect that the new data I'm labeling won't be redundant? To my knowledge, if a datapoint is labeled automatically, unless novel data is generated like in this paper I don't see how the data that can be labeled will be particularly useful, as it will admit no new patterns to learn from, as otherwise the data wouldn't be labeled with a high confidence. Is the point here to make sure the labeled dataset you start with is sufficiently large that it covers all the kinds of patterns you expect from the data for the model to need to be able to identify?

$\endgroup$

1 Answer 1

0
$\begingroup$

If you are confident that your pseudo-labels are correct, then you can expect that the new data you label will not be redundant. however, If you want to avoid having redundant data in your labeled dataset, then you need to make sure that it is large enough to cover all the relevant patterns.

One way to avoid creating redundant data when using data augmentation is to use a technique called transfer learning. This involves training a model on one dataset and then transferring the learned knowledge to another related dataset. This can be done by fine-tuning the model on the new dataset or by using the model to extract features from the new dataset that can be used to train a new model. Transfer learning can be used to reduce the amount of data needed to train a new model, and can also improve the performance of the new model.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .