Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail Learning
Authors: Xiaoling Zhou, Ou Wu, Nan Yang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across diverse scenarios susceptible to class and attribute imbalances showcase the state-of-the-art performance of Meta-CALA. Furthermore, while Heuristic-CALA exhibits inferior performance compared to Meta-CALA, it incurs only negligible additional training time compared to the Cross-Entropy loss, yet surpasses existing methods by a significant margin. We conduct extensive experiments across three learning scenarios susceptible to class and attribute imbalances: LT learning, GLT learning, and subpopulation shift learn-ing. |
| Researcher Affiliation | Academia | 1Center for Applied Mathematics, Tianjin University, China 2 National Engineering Research Center for Software Engineering, Peking University, China 3 HIAS, University of Chinese Academy of Sciences, Hangzhou, China EMAIL |
| Pseudocode | No | The paper describes the methods using mathematical equations and text, such as: "ˆ W (t) W (t) η1 1 n i=1 W ℓCALA f(xi), yi; δ(t) i , (9)" and "Ω(t+1) Ω(t) η2 1 m i=1 ΩℓCE f ˆ W (xmeta i ), ymeta i , (10)". It does not contain a dedicated pseudocode or algorithm block. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Three LT benchmarks are evaluated: CIFAR-LT (Cui et al. 2019), Places-LT (Liu et al. 2019), and i Naturalist (i Nat) 2018. Three subpopulation shift datasets are adopted: CMNIST (Arjovsky et al. 2019), Waterbirds (Sagawa et al. 2020), and Celeb A (Liu et al. 2015), each exhibiting significant attribute imbalances within classes. We consider two GLT benchmarks, Image Net-GLT and MSCOCO-GLT (Tang et al. 2022). |
| Dataset Splits | No | The paper mentions using standard datasets like CIFAR-LT, Places-LT, i Nat 2018, CMNIST, Waterbirds, Celeb A, Image Net-GLT, and MSCOCO-GLT. It also details the construction of a metadata set, e.g., 'The metadata size is 3,000 for CIFAR-LT. For i Nat 2018 and Places-LT, one image is selected per class and group to construct the metadata.' However, it does not explicitly provide the specific training, validation, and test splits (e.g., percentages or exact counts) used for the main model evaluation, nor does it explicitly cite predefined splits for these datasets within the text. |
| Hardware Specification | No | For CIFAR-LT, we primarily follow Cao et al. (2019) and train all models with a Res Net-32 (He et al. 2016) backbone on a single GPU... |
| Software Dependencies | No | The paper mentions using an SGD optimizer and Adam optimizer, and various neural network architectures (ResNet-32, ResNet-152, ResNet-50, ResNeXt-50). However, it does not specify any software libraries or packages with their version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For all experiments, we utilize the SGD optimizer with a momentum of 0.9. For CIFAR-LT, we primarily follow Cao et al. (2019) and train all models with a Res Net-32 (He et al. 2016) backbone on a single GPU, employing a multistep learning rate schedule that reduces the learning rate by a factor of 0.01 at the 160th and 180th epochs. For Places-LT and i Nat 2018, we mainly follow Kang et al. (2020) and use the cosine learning rate schedule (Loshchilov and Hutter 2016) to train the Res Net152 and Res Net-50 backbones, respectively. For the hyperparameters in CALA, the neighborhood size K is selected from {20, 40, 60, 80, 100} for all experiments unless noted. τ1 and τ2 are set to 1.5 and 1, respectively. In Meta-CALA, the adjustment network is optimized using Adam with an initial learning rate of 1 10 3. |