Iterative label cleaning for transductive and semi-supervised few-shot learning

Given a set of completely unlabeled images (left) and very few or just one clean image per class with one image-level label per image and no other information (right), we learn

This motivates us to study a problem where the representation is obtained from a classifier pre-trained on a large-scale dataset of a different domain, assuming no access to

Concretely, we prove three multiclass versions and one multi-label version of the C-bound, a bound over the risk of the majority vote, based on generalizations of the notion of

Face recognition from a single image per person is a diffi- cult problem [1] because all state-of-the-art face recognizers rely on a large number of training data, which are

Figure 2 and Figure 3 further show that semi- supervised training using label propagation on transformed representations (‘h1’ and ‘h2’) per- forms better than the ‘Baseline’

TAG_2005 TAG_2006 TAG_2007 TAG_agdetection TAG_algorithms TAG_amperometry TAG_analysis TAG_and TAG_annotation TAG_antibody TAG_apob TAG_architecture TAG_article

We immediately notice that at both adaptation speeds our models are capable of surpassing the all-data model 11 in terms of accuracy on the new words, despite having seen fewer than

In the following set of experiments with the practical FOM and CG algorithms, we assume that the products Ap can be computed in double, single or half precision (with