DIFAIR: Towards learning differenciated and interpretable representations

Quentin Christoffel, Aline Deruyver, Ali Ayadi, Anne Jeannin-Girardon

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: OSR, open set recognition, representation learning, interpretability, feature visualization, computer vision

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We present separated class representations and interpretable representations, and how it can be beneficial for Open Set Recognition.

Abstract: Neural network classifiers are generally trained to differentiate between the same classes during training and testing. In order to prevent incorrect predictions, when an input image contains a class that was not part of the training set, it should be detected. The process of detection of \`\`unknown'' classes is called Open Set Recogniton (OSR). Given that a neural network extracts a representation (a feature vector) describing an image, its capacity to detect the presence of a class in an image, through the recognition of specific features, should also imply the ability to detect the absence of a \`\`known'' class, through the absence of those features in the representation. In this article, we present DIFAIR, a novel approach introducing the key characteristics that a feature representation should exhibit to ensure: (i) class separability, through predefined class positions in the representation space; and (ii) interpretability by associating each dimension of the representation with a class. We present a loss function to optimize a model, in a supervised way, in order to produce the proposed representation. Our approach assumes that unknown classes should share only a limited number of features with known classes and therefore we evaluate its performance in OSR. Finally, we visually inspect learned representations to identify the flaws of our loss function and present directions for future improvement.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5353