A Bag of Tricks for Few-Shot Class-Incremental Learning
Authors: Shuvendu Roy, Chunjong Park, Aldi Fahrezi, Ali Etemad
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on three benchmark datasets, CIFAR-100, CUB-200, and mini IMage Net, to evaluate the impact of our proposed framework. Our detailed analysis shows that our approach substantially improves both stability and adaptability, establishing a new state-of-the-art by outperforming prior works in the area. |
| Researcher Affiliation | Collaboration | Shuvendu Roy1,2 , Chunjong Park1, Aldi Fahrezi1, Ali Etemad1,2 1Google Research 2Queen s University, Canada EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper includes mathematical equations for loss functions and masks but does not present any structured pseudocode blocks or algorithms. |
| Open Source Code | No | The paper does not provide a direct link to a source code repository, nor does it explicitly state that the code for the described methodology will be made publicly available or is included in supplementary materials. |
| Open Datasets | Yes | Following the established protocol in the FSCIL literature, we conduct our experiments on three popular datasets: CIFAR100 (Krizhevsky et al., 2009), mini Image Net (Russakovsky et al., 2015) and CUB200 (Wah et al., 2011). |
| Dataset Splits | Yes | Specifically, for CIFAR-100 and mini Image Net, we use 60 classes for the base session and 40 classes for the incremental sessions. The incremental learning experiments are conducted on a 5-way, 5-shot setting. In the case of CUB-200, we allocate 100 classes for the base session and another 100 classes for the incremental sessions, each containing ten classes (10-way, 5-shot). |
| Hardware Specification | Yes | In Table 6, we discuss the time complexity of our framework using a single Nvidia RTX 2080 GPU in comparison to SAVC (Song et al., 2023). |
| Software Dependencies | No | The paper mentions using a Res Net-18 encoder and an SGD optimizer but does not specify software versions for programming languages, libraries, or frameworks like Python, PyTorch, TensorFlow, or CUDA. |
| Experiment Setup | Yes | We train the model with an SGD optimizer, a momentum of 0.9, and a batch size of 64. The learning rate is set to 0.1 for CIFAR-100 and mini Image Net and 0.001 for CUB-200. For all experiments, the model is trained on an Nvidia RTX 2080 GPU. Our findings from this study show that the best results for CIFAR-100, CUB-200 and mini Image Net datasets are obtained for training 400, 80, and 80 epochs, respectively. |