Summary of the Challenge

The challenge tackles the Class-Incremental with Repetition (CIR) scenario using unlabelled data, where previously observed classes re-appear with varying repetition and not all classes are present in each experience. Each of the three challenge scenarios consists of 50 experiences with training sessions using both labelled and unlabelled samples, which become unavailable after training. Future experiences may include seen, unseen, or distractor classes. Participants must develop strategies that achieve high accuracy on a balanced evaluation test set after training on the entire stream, with the same algorithm applied across all scenarios. 

Videos of the workshop talk and the solutions TBA

Top Performing Teams

1st place: Team NJUST-KMG

Description

Team NJUST-KMG uses Winning SubNetworks, a parameter isolation method, combined with unsupervised contrastive learning on unlabeled data to learn robust features. They also apply the FixMatch technique with a high threshold across three scenarios to enhance model performance. During inference, they average prediction logits across different tasks to leverage the diverse features learned at each stage.

Team Members

Sishun Pan, Tingmin Li, Yang Yang

Affiliation

Nanjing University of Science and Technology

2nd place: Team SNUMPR

Description

Team SNUMPR enhances the ER-ACE method by incorporating unlabeled data to encourage both plasticity and stability. They apply adaptively weighted self-supervised learning (SSL) using rotation prediction and knowledge distillation based on feature and logit vectors from previous tasks. They utilize a memory buffer to store 200 exemplars, each containing a feature vector, logit vector, and label, replaying these to mimic stored logit vectors and labels. During inference, they use model ensembling, averaging the outputs of the current and previous models to achieve the final prediction logits.

Team Members

Taeheon Kim, San Kim, Dongjae Jeon, Minhyuk Seo, Wonje Jeong, Jonghyun Choi

Affiliation

Seoul National University, Yonsei University

3rd place: Team CytunAI

Description

Team CytunAI employs a Frequency-Aware (FA) storage policy to address class imbalance in continual learning by dynamically adjusting buffer slots based on the repetition frequency of each class. This approach helps mitigate the imbalance in the number of instances per class in the data stream. They save the average feature before the fully connected (FC) layer for each class and compute the loss of this feature on both the new and old models' FC layer outputs. Additionally, for the current input, they calculate the loss of features before the FC layer and the loss of outputs from both the new and old models, facilitating effective continual learning.

Team Members

Chengkun Ling, Weiwei Zhou 

Affiliation

China Telecom

4th place: Team PM_EK

Description

Team PM_EK extends the baseline Learning without Forgetting (LWF) strategy by applying it to both labeled and unlabeled data streams. They generate class prototypes from labeled data and assign pseudo-labels to unlabeled data based on feature similarity to these prototypes, avoiding pseudo-labeling when similarity is below a threshold. This method helps handle unseen or distractor classes. They also apply less forgetful learning (LFL) strategies and employ learning rate scheduling and early stopping to prevent overfitting, with a training epoch limit set to 15.

Team Members

Panagiota Moraiti, Efstathios Karypidis

Affiliation

Democritus University of Thrace (DUTh), National Technical University Of Athens (NTUA), Athena Research Center - Archimedes Unit