Rethinking the Bias of Foundation Model under Long-tailed Distribution

Authors: Jiahao Chen, Bin Qin, Jiangmeng Li, Hao Chen, Bing Su

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we examine how such imbalances from pre-training affect long-tailed downstream tasks. Specifically, we find the imbalance biases inherited in foundation models on downstream tasks as parameter imbalance and data imbalance. ... We achieve at least 1.5%, 1.5%, 2.0% performance gains on Image Net-LT (Deng et al., 2009), Places365LT (Liu et al., 2019), and i Naturalist2018 (Van Horn et al., 2018) compared with state-of-the-art methods.
Researcher Affiliation Academia 1Gaoling School of Artificial Intelligence, Renmin University of China 2Beijing Key Laboratory of Research on Large Models and Intelligent Governance 3Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE 4Institute of Software Chinese Academy of Sciences 5University of Chinese Academy of Sciences 6Electrical and Computer Engineering, Carnegie Mellon University. Correspondence to: Bing Su <EMAIL>.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing code or a link to a code repository.
Open Datasets Yes We achieve at least 1.5%, 1.5%, 2.0% performance gains on Image Net-LT (Deng et al., 2009), Places365LT (Liu et al., 2019), and i Naturalist2018 (Van Horn et al., 2018) compared with state-of-the-art methods.
Dataset Splits No Following OLTR (Liu et al., 2019), we split the classes into three groups named D-Many , D-Medium , and D-Few relying on the number of samples. Similarly, for parameter imbalance, we split the classes into three groups named P-Many , P-Medium , and P-Few relying on b PP (Y ). More details are in the Appendix Sec. A. ... Additionally, in Tab.11, we provide results highlighting the performance under parameter imbalance.
Hardware Specification Yes For training resources, all experiments are conducted on Intel(R) Xeon(R) Gold 5318Y CPU @ 2.10GHz with a single RTX A40 GPU. Normally, a GPU with 24GB of memory is sufficient for the reproduction.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python version, PyTorch/TensorFlow version, CUDA version).
Experiment Setup Yes We present the details about the hyper-parameters of our experiments on different datasets in Tab. 9, where lr, epochs denote the initial learning rate and training epochs, respectively. We denote batch size in Tab. 9 as the training batch size during the fine-tuning phase. ... The learning rate, number of epochs, and parameter initialization strategies follows (Shi et al., 2024).