Semi-Supervised Image Classification (Cold Start)

1 papers with code • 8 benchmarks • 1 datasets

This is the same as the semi-supervised image classification task, with the key difference being that the labelled subset chosen needs to be selection in a class agnostic manner. This means that the standard practice in semi-supervised learning of using a random class stratified sample is "cheating" in this case, as class information is required for the whole dataset for this to be done. Rather, this challenge requires a smart cold-start or unsupervised selective labelling strategy to identify images that are most informative and result in the best performing models.

Datasets


Most implemented papers

Unsupervised Selective Labeling for More Effective Semi-Supervised Learning

TonyLianLong/UnsupervisedSelectiveLabeling 6 Oct 2021

Intuitively, no matter what the downstream task is, instances to be labeled must be representative and diverse: The former would facilitate label propagation to unlabeled data, whereas the latter would ensure coverage of the entire dataset.