# TACLE: Task and Class-aware Exemplar-free Semi-supervised Class Incremental Learning

**Authors:** Jayateja Kalla, Rohit Kumar, Soma Biswas  
**arXiv:** [2407.08041](https://arxiv.org/abs/2407.08041)  
**Date:** 2024-07-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.08041v1/

## Key points

- TACLE (TAsk and CLass-awarE) framework proposed for exemplar-free semi-supervised class incremental learning
- Enables learning new classes at each task using labeled and unlabeled data without access to exemplars from previous classes
- Leverages pre-trained models, task-adaptive threshold, and class-aware weighted cross-entropy loss
- Utilizes unlabeled data for classifier alignment to enhance model performance
- Experimental validation on CIFAR10, CIFAR100, and ImageNet-Subset100 datasets demonstrates effectiveness
- Implementation details include estimating mean and variance, calculating alignment loss for classifiers, following experimental protocols
- Significant improvements in accuracy across all datasets with TACLE achieving 92.35% ± 0.06 accuracy on CIFAR100
- Demonstrates robust performance in exemplar-free semi-supervised class incremental learning scenarios

## Summary

The TACLE (TAsk and CLass-awarE) framework is proposed to address the challenging problem of exemplar-free semi-supervised class incremental learning. This approach enables the model to learn new classes at each task using both labeled and unlabeled data without access to exemplars from previous classes. The framework leverages pre-trained models and introduces a task-adaptive threshold to maximize the utilization of available unlabeled data as learning progresses. To handle under-represented classes within tasks, a class-aware weighted cross-entropy loss is also introduced. Additionally, the framework utilizes unlabeled data for classifier alignment in order to enhance model performance. Experimental validation on benchmark datasets including CIFAR10, CIFAR100, and ImageNet-Subset100 demonstrates the effectiveness of TACLE. The framework shows promising results even with imbalanced unlabeled data and extreme scenarios with only one labeled example per class. Implementation details include estimating mean and variance using dataset information, calculating alignment loss for classifiers, and following experimental protocols. The datasets used in evaluation consist of CIFAR10 with 5 tasks incrementally learning 2 classes per task, CIFAR100 with 10 tasks introducing 10 new classes per task, and ImageNet-Subset100 containing 100 classes. Results showcase significant improvements in accuracy across all datasets, with TACLE achieving an accuracy of 92.35% ± 0.06 on CIFAR100. Overall, the proposed framework demonstrates robust performance in exemplar-free semi-supervised class incremental learning scenarios.