Co–TES: Learning Noisy Labels with a Co-Teaching Exchange Student Method

Jul 6, 2026

Upload by Jiven

Machine-learning performance depends on both model structure and dataset quality, yet large-scale labeled data is costly and often contains noise. Learning with Noisy Labels targets robust training on mixtures of clean and noisy samples. The proposed Co-TES algorithm improves co-teaching by using models with different architectures, enabling each iteration’s data selection to capture complementary features. Experiments indicate faster gains in early-to-mid stages and stable learning compared with traditional loss-based training.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Introduction

Learning with Noisy Labels and robust training

Multi-network learning and its limitations

Related work

Designing robust components

Method and experimental setup

Proposed Co-TES approach

Experimental environment, datasets, and model structure

Results and conclusions

Experimental results analysis

Conclusions and future work

Pattern Recognition Letters 182 (2024) 17–23 | Co–TES: Learning noisy labels with a Co-Teaching Exchange Student method Chan Ho Shin, Seong-jun Oh ∗<br>Korea University, Seoul, Republic of Korea | | | | | --- | --- | --- | --- | | A R T I C L E I N F O | | A B S T R A C T | | | Editor: Jose Ruiz-Shulcloper | | The performance of a machine-learning model is influenced by two main factors: the structure of the model, and the quality of the dataset it processes. As high-quality labeled data in substantial size is often difficult to obtain, there are ongoing efforts to develop machine learning algorithms that are robust with noisy datasets. Among these algorithms, multi-network learning utilizes learning from a noisy dataset by the selection and filtering of samples through multiple learning networks. We propose an improved co-teaching algorithm termed Co-TES that leverages different models with various architectures. Co-TES extracts different features from each iteration of data selection and makes the model more robust with the same quality dataset. Numerical results show that the proposed method can lead to faster performance gains in the early to mid-range. | | | Keywords:<br>Learning with noisy labels<br>Co-teaching<br>Multi-network learning | | | | 1. Introduction Deep learning models with large-scale datasets have brought tremendous success to AI research, in computer vision [1], voice synthesis [2], and natural language processing [3], among other fields. While the amount of available raw data has increased, it remains difficult and expensive to label it manually. While dataset labeling has been performed by crowdsourcing [4], crawling [5], and other low-cost or automated solutions, these techniques often result in a lower-quality dataset that inevitably leads to poor performance. Recent research [6,7] has shown that the performance of a deep learning model depends both on the structure of the model, and on the quality of the dataset it processes. Consequently, the quality of the dataset is critical for model training, and obtaining high-quality data is a challenging issue [8,9]. Learning with Noisy Labels (LNL) aims to achieve robust learning on a dataset that consists of clean and noisy data. Robust learning can be conducted by modifying model structures [4,10,11], regularization [12–15], and loss functions [16–18]. Using the abovementioned modifications, several sample selection methods have been proposed [19–25]. They can be grouped into multi-round learning [19, 20] and multi-network learning [21–25]. In particular, multi-network learning has received much attention, since it can prevent bias with only a small number of clean samples selected at the initial training. The above-mentioned multi-network learning methods are limited to networks of the same architecture and size; therefore, the performance of the model should be greatly affected by the weights of the initial network. To the best of our knowledge, no co-teaching algorithm has previously been proposed that uses models with different architectures, we being the first to do so. Since different neural networks extract different features from different perspectives, many researchers have tried to develop models that perform better in their fields, rather than models that perform well generally. We too focus on the fact that different architectures can extract different features, and provide fresh, more accurate perspectives. In this paper, we propose a method termed Co-TES that mutually trains models with a combination of different architectures. By using models with different architectures to select training data and extract different features, Co-TES achieves robust learning that is less affected by the quality of the dataset. We also focus only on text data, since the same text embedding allows easy comparison of different combinations of networks with varying structures. Co-TES showed a faster initial performance improvement than other mul

FAQs

What problem does Co-TES address?

Co-TES addresses robust learning when training data contains noisy labels, where low-quality labeled data can harm model performance.

How does Co-TES differ from standard multi-network or co-teaching methods?

Co-TES uses models with different architectures to select training data and extract features, aiming to reduce sensitivity to the initial network’s weights.

What performance behavior is reported for the proposed method?

Numerical results show faster performance improvement in early to mid training and stable learning behavior compared with traditional loss-based methodology.

Jul 6, 2026

Upload by Jiven