Making Reliable and Flexible Decisions in Long-tailed Classification

Authors: Bolian Li, Ruqi Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental In empirical evaluation, we design a new metric, False Head Rate, to quantify tail-sensitivity risk, along with comprehensive experiments on multiple real-world tasks, including large-scale image classification and uncertainty quantification, to demonstrate the reliability and flexibility of our method. We conduct comprehensive experiments to demonstrate that RF-DLC significantly improves decision-making while maintaining or improving traditional metrics such as accuracy and calibration.
Researcher Affiliation Academia Bolian Li EMAIL Ruqi Zhang EMAIL Department of Computer Science, Purdue University West Lafayette, IN 47907, USA
Pseudocode Yes The proposed RF-DLC is summarized in Algorithm 1.
Open Source Code Yes 1https://github.com/lblaoke/RF-DLC.
Open Datasets Yes We use CIFAR10/100-LT (Cui et al., 2019), Image Net-LT (Liu et al., 2019b), and i Naturalist (Van Horn et al., 2018) as the long-tailed datasets... We also conduct a long-tailed medical image classification experiments in Table 5, where our RF-DLC successfully recognize different disease types and outperforms compared baselines. The experiments are based on Res Net32, and the number of particles is set to 3. The Derma MNIST dataset (Yang et al., 2023) is originally an imbalanced classification dataset with the imbalance ratio to be around 60.
Dataset Splits Yes In long-tailed categorical data, the training and testing sets follow different distributions... Classes are equally split into three class regions (head, med and tail). For example, there are 33, 33 and 34 classes respectively in the head, med and tail regions of CIFAR100-LT.
Hardware Specification Yes We run all experiments on an NVIDIA RTX A6000 GPU (49 GB) and do not need multiple GPUs for one model.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers.
Experiment Setup Yes Table 10: Hyper-parameter configurations. Dataset Base Model Optimizer Batch Size Learning Rate Training Epochs Discrepancy Ratio λ τ α CIFAR10-LT Res Net32 SGD 128 0.1 200 linear 5e-4 40 0.002 CIFAR100-LT Res Net32 SGD 128 0.1 200 linear 5e-4 40 0.3 Image Net-LT Res Net50 SGD 256 0.1 100 linear 2e-4 20 50 i Naturalist Slim Res Net50 SGD 512 0.2 100 linear 2e-4 20 100