Novel Class Discovery Without Forgetting

Abstract

Humans possess an innate ability to identify and differentiate instances that they are not familiar with, by leveraging and adapting the knowledge that they have acquired so far. Importantly, they achieve this without deteriorating the performance on their earlier learning. Inspired by this, we identify and formulate a new, pragmatic problem setting of NCDwF: Novel Class Discovery without Forgetting, which tasks a machine learning model to incrementally discover novel categories of instances from unlabeled data, while maintaining its performance on the previously seen categories. We propose 1) a method to generate pseudolatent representations which act as a proxy for (no longer available) labeled data, thereby alleviating forgetting, 2) a mutual-information based regularizer which enhances unsupervised discovery of novel classes, and 3) a simple Known Class Identifier which aids generalized inference when the testing data contains instances form both seen and unseen categories. We introduce experimental protocols based on CIFAR-10, CIFAR-100 and ImageNet-1000 to measure the trade-off between knowledge retention and novel class discovery. Our extensive evaluations reveal that existing models catastrophically forget previously seen categories while identifying novel categories, while our method is able to effectively balance between the competing objectives. We hope our work will attract further research into this newly identified pragmatic problem setting.

Novel Class Discovery: an Introduction and Key Concepts

“…Novel Class Discovery without Forgetting (NCDwF) [45] is another domain that relaxes some of the assumptions behind NCD. In NCDwF, D l and D u are not available simultaneously.…”

“…Finally, [45] introduces the name NCDwF. To avoid the forgetting, it proposes a method to generate synthetic samples that are representative of each known class and act as a proxy for the no longer available labeled data.…”

A cache cooperation management for wireless

“…Keeping in mind the data regulatory practices, the NCD community has been paying more attention to the problem of Incremental Novel Class Discovery (iNCD) [31] where the access to the labelled (or base) dataset is absent during the discovery stage. Unlike iNCD, Roy et al [37] and Joseph et al [26] investigate a more realistic yet challenging setting known as Class-incremental Novel Class Discovery (class-iNCD), where task-id information is not available during inference.…”

“…However, no methods to date have explored the Multistep Class-incremental Novel Class Discovery (MSc-iNCD), where the goal is to continuously discover novel classes in a sequence of unlabelled data sets over multiple steps, rather than the few steps in the class-iNCD literature (2 steps in [26] and 1 step in [26]). In this work, we study the problem of MSc-iNCD (see Fig.…”

“…Due to storage and privacy reasons the access to the labelled dataset can be revoked, which makes NCD a very challenging problem. To address this, some very recent Class-incremental Novel Class Discovery (class-iNCD) methods [37,26] have attempted to address NCD from the lens of continual learning, by not accessing the labelled dataset when learning new classes (see Fig. 1c).…”