Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples
Authors: Suqin Yuan, Lei Feng, Bo Han, Tongliang Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on CIFAR (Krizhevsky et al., 2009), Web Vision (Li et al., 2017), and full Image Net-1k (Deng et al., 2009) datasets with different types and levels of label noise. The results demonstrate that our proposed method consistently outperforms state-of-the-art sample selection methods across various computer vision tasks. |
| Researcher Affiliation | Academia | Suqin Yuan1 Lei Feng2 Bo Han3 Tongliang Liu1 1 Sydney AI Centre, The University of Sydney 2 Southeast University 3 Hong Kong Baptist University |
| Pseudocode | Yes | We formalize the complete Early Cutting procedure in Algorithm 1. Algorithm 1 Iterative Sample Selection with Early Cutting Require: Training data D0; Number of iterations Irate; Early cutting rate γ; Thresholds δ, τ, ϵ Ensure: Trained model parameters θ 1: for i = 1 to Irate do 2: 1. Base Sample Selection: 3: Compute learning times LTj for all (xj, yj) Di 1 using Eq. (1) 4: Select the initial early-learned subset Ds based on smallest learning times. 5: 2. Early Cutting: 6: Select candidate subset D s from Ds (e.g., the 1 γ proportion with the earliest learning times). 7: Compute loss Lj, confidence cj, and gradient norm gj for all samples in D s. 8: Identify the set of MEEs in D s according to Definition 1: 9: MEEs = j D s | (Lj > δ) (cj > τ) (gj < ϵ) 10: Create the refined subset by removing the identified MEEs: Ds refined Ds \ MEEs. 11: Update the training data for the next iteration: Di Ds refined. 12: end for 13: Final Training Phase: Train a model from scratch on the final refined set DIrate until convergence. 14: return Trained model parameters θ . |
| Open Source Code | Yes | Our implementation can be found at https://github.com/tmllab/2025_Neur IPS_MEE. |
| Open Datasets | Yes | We conduct extensive experiments on CIFAR (Krizhevsky et al., 2009), Web Vision (Li et al., 2017), and full Image Net-1k (Deng et al., 2009) datasets with different types and levels of label noise. |
| Dataset Splits | Yes | We split 10% noisy trianing data for validation. For the CIFAR-N task, we utilized the provided noisy labels. |
| Hardware Specification | Yes | Table 8 provides a comparative overview of total training hours and performance for various methods and our Early Cutting variants on CIFAR-100 with Res Net-34, tested on a single NVIDIA RTX 4090 GPU. |
| Software Dependencies | No | Training was performed using SGD (Robbins and Monro, 1951) with a momentum (Rumelhart et al., 1986) of 0.9 and a weight decay (Krogh and Hertz, 1991) of 5 10 4. The initial learning rate was set to 0.1 and decayed using a cosine annealing schedule. |
| Experiment Setup | Yes | Training was performed using Stochastic Gradient Descent (SGD) with a momentum of 0.9 and a weight decay of 5 10 4. The initial learning rate was set to 0.1 and decayed using a cosine annealing schedule without restarts, decreasing to 1 10 5 over the course of training. The number of training epochs was set to 300 for CIFAR, 200 for Web Vision, and 150 for full Image Net-1K experiments. Batch sizes were set to 32 for CIFAR datasets and Web Vision, and 256 for Image Net-1K. |