What does "class-level discriminative feature representation" mean in the paper "Semi-Supervised Deep Learning with Memory"?

Question

I am reading the paper Semi-Supervised Deep Learning with Memory (2018) by Yanbei Chen et al. The topic is the classification of images using semi-supervised learning. The authors use a term on page 2 in the middle of the page that I am not familiar with. They write:

The key to our framework design is two-aspect: (1) the class-level discriminative feature representation and the network inference uncertainty are gradually accumulated in an external memory module; (2) this memorised information is utilised to assimilate the newly incoming image samples on-the-fly and generate an informative unsupervised memory loss to guide the network learning jointly with the supervised classification loss

I am not sure what the term discriminative feature representation means.

I know that a discriminative model determines the decision boundary between the classes, and examples include: Logistic Regression (LR), Support Vector Machine (SVM), conditional random fields (CRFs) and others.

Moreover, I know that, in machine learning, feature learning or representation learning is a set of techniques that allows a system to automatically discover the representations needed for feature detection or classification from raw data.

Any insights on the definition of this term much appreciated.

Answer 1

Discriminative models give the probability of an element in the feature space $x \in X$ belonging to a class $y \in Y$, i.e. $p(Y|X)$, where $Y$ is the set of classes in a classification problem.

The discriminative feature representation in this context means the feature map/s which are the output/s of the convolutional layers of the backbone (the convolutional layers of the neural network) which (presumably) differ on the basis of their class, and can be used to discriminate which class the original image belongs to, in the case of image classification.