Cross-Domain Empirical Risk Minimization for Unbiased Long-Tailed Classification

Authors: Beier Zhu, Yulei Niu, Xian-Sheng Hua, Hanwang Zhang3589-3597

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on several long-tail classification benchmark datasets: CIFAR100-LT (Cui et al. 2019), Places365-LT (Liu et al. 2019), Image Net-LT (Liu et al. 2019), and i Naturalist 2018 (Van Horn et al. 2018). Experimental results show that x ERM outperforms previous stateof-the-arts on both long-tailed and balanced test sets, which demonstrates that the performance gain is not from catering to the tail. Further qualitative studies show that the x ERM helps with better feature representation.
Researcher Affiliation Collaboration Beier Zhu1, Yulei Niu1*, Xian-Sheng Hua2, Hanwang Zhang1 1Nanyang Technological University 2Damo Academy, Alibaba Group
Pseudocode Yes Figure 2: The Algorithm: x ERM
Open Source Code No The paper mentions that "The implementations of the two models are open" but does not provide a specific link or explicit statement about releasing the code for x ERM itself.
Open Datasets Yes We conducted experiments on four long-tailed classification datasets: CIFAR100-LT (Cui et al. 2019), Places365-LT (Liu et al. 2019), Image Net-LT (Liu et al. 2019), and i Naturalist 2018 (Van Horn et al. 2018).
Dataset Splits Yes We divided the test set of CIFAR100-LT-IB-100 and Image Net LT into three subsets according to the number of samples in each class: many-shot (categories with >100 images), medium-shot (categories with 20 100 images), and few-shot (categories with <20 images). ... We established the long-tailed test splits by downsampling the original well-balanced test set with various imbalanced ratios, which is the same as the training set construction.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes We set the scaling parameter γ in Eq. (9) to 2 for CIFAR100-LT, 5 for Places365-LT and 1.5 for other datasets. ... All networks were trained for 200 epochs on CIFAR100-LT, 30 epochs on Places365-LT, and 90 epochs on Image Net-LT and i Naturalist18.