Active and Semi-Supervised machine learning: Nov 16–Dec 4 | by Olga Petrova | Dec, 2020

[ad_1]


Active learning is, of course, not the only way to deal with the problem of not having enough labelled data. Semi-supervised learning (of which active learning can be viewed as an example) has a wide variety of other methods, where unlabelled data is used alongside the labelled training samples for the common goal. Then there is self-supervised learning: technically supervised (each data point is assigned a target), it requires no manual annotation on our part.

As the cheeky title of the preprint above suggests, the authors found that when semi-supervised and self-supervised approaches join forces, the benefit of adding active learning into the mix can be marginal, at best. However, keep in mind that what is true for CIFAR10 that the authors chose, might have little bearing on the results you obtain for your dataset. (Of course, the same goes for most other active learning studies, including the more optimistic ones.)

Read More …

[ad_2]


Write a comment