A Unified Generalization Analysis of Re-Weighting and Logit-Adjustment for Imbalanced Learning

Authors: Zitai Wang, Qianqian Xu, Zhiyong Yang, Yuan He, Xiaochun Cao, Qingming Huang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, the empirical results on benchmark datasets not only validate the theoretical results but also demonstrate the effectiveness of the proposed method.
Researcher Affiliation Collaboration 1 SKLOIS, Institute of Information Engineering, CAS 2 School of Cyber Security, University of Chinese Academy of Sciences 3 Key Lab. of Intelligent Information Processing, Institute of Computing Tech., CAS 4 School of Computer Science and Tech., University of Chinese Academy of Sciences 5 Alibaba Group 6 School of Cyber Science and Tech., Shenzhen Campus of Sun Yat-sen University 7 BDKM, University of Chinese Academy of Sciences
Pseudocode Yes Algorithm 1: Principled Learning Algorithm induced by the Theoretical Insights
Open Source Code No The paper does not provide a link to its source code or explicitly state its release.
Open Datasets Yes Datasets. We conduct the experiments on four popular benchmark datasets for imbalanced learning. (a) CIFAR-10 and CIFAR-100: Following the protocol in [26, 11, 12]... (b) Image Net-LT and i Naturalist: We use the long-tailed version of the Image Net dataset2 [2] proposed by [27], and i Naturalist3 [5] is a real-world long-tailed dataset.
Dataset Splits Yes Datasets. We conduct the experiments on four popular benchmark datasets for imbalanced learning. (a) CIFAR-10 and CIFAR-100: Following the protocol in [26, 11, 12]... (b) Image Net-LT and i Naturalist: We use the long-tailed version of the Image Net dataset2 [2] proposed by [27], and i Naturalist3 [5] is a real-world long-tailed dataset.
Hardware Specification No The paper does not specify the hardware used for running the experiments.
Software Dependencies No The paper mentions 'Pytorch' [51], 'numpy' [52], 'scikit-learn' [53], and 'Torchvision' [54] but does not provide specific version numbers for these software dependencies as used in their experiments.
Experiment Setup Yes We tune the hyperparameter ν, and the other hyperparameters follow those used in the baselines. In addition, we incorporate the Sharpness-Aware Minimization (SAM) technique [28]... Algorithm 1 details: if t < T0 then Set α = 1, βy, y Adjust logits during the initial phase ... else Set αy π ν y , βy = 1, y, ν > 0 TLA and ADRW ... Θ Θ η ΘL(f, B) One SGD step Optional: anneal the learning rate η. Required when t = T0. ... Figure 3: (a) Training accuracy of CE+DRW (T0 = 160)