CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning

Authors: Qiwei Li, Jiahuan Zhou

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on various datasets demonstrate that our proposed CAPrompt outperforms state-of-the-art methods by 2%-3%.
Researcher Affiliation Academia Qiwei Li, Jiahuan Zhou* Wangxuan Institute of Computer Technology, Peking University, Beijing 100871, China EMAIL
Pseudocode Yes Algorithm 1: Cyclic Prompt Aggregation algorithm Training: Input: Data stream D = {Dt}T t=1, a pre-trained Vi T f( ). Output: Prompts ϕ1, ..., ϕT and classification head W.
Open Source Code Yes Code https://github.com/zhoujiahuan1991/AAAI2025CAPrompt
Open Datasets Yes We follow existing works (Wang et al. 2024a) to evaluate our proposed method on three public datasets, CIFAR-100 (Krizhevsky, Hinton et al. 2009), Image Net R (Hendrycks et al. 2021), and CUB200 (Wah et al. 2011).
Dataset Splits No The paper states: 'These datasets are splited into 10 tasks with disjoint classes for incremental learning.' However, it does not provide specific percentages or counts for training, validation, or test splits within these tasks, nor does it explicitly reference standard splits for these datasets for reproducibility beyond the initial task splitting.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only mentions 'Image Net21K pre-trained Vi T-B/16 as the backbone', which refers to a model architecture rather than the underlying hardware.
Software Dependencies No The paper mentions using an 'Adam optimizer' and 'Prefix Tuning Prompts' as techniques but does not provide specific version numbers for any software libraries, programming languages (e.g., Python), or frameworks (e.g., PyTorch, TensorFlow, CUDA) that would be needed for replication.
Experiment Setup Yes The parameters are optimized by an Adam optimizer with an initial learning rate of 3e-3 and a batch size of 24. Prefix Tuning Prompts (Wang et al. 2022c), with the prompt length Lp = 10 are inserted into all layers for CIFAR-100, CUB200, and into the first nine layers for Image Net-R with Lp = 20. ... The weighting parameters of different losses are α = 5 and β = 0.2. The cyclic number num is set to 2 without special explanations for a similar computation cost compared to existing methods.