Identifying Mislabeled Images in Supervised Learning Utilizing Autoencoder
Autoencoder-based outlier detection with DBSCAN removes over 67% mislabeled images, improving dataset quality.
Key Findings
Methodology
This paper introduces a framework combining convolutional autoencoders and density-based clustering (DBSCAN) to identify mislabeled samples in training datasets. Images are encoded into a low-dimensional latent space where similar features cluster together. Using Rahmah’s algorithm, optimal DBSCAN parameters are adaptively determined for each class. Outliers, identified as samples with features deviating from the main cluster, are considered mislabeled and removed. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 datasets show detection rates exceeding 67%, significantly enhancing data quality for supervised learning. The approach leverages unsupervised feature extraction and density-based anomaly detection, requiring no modifications to the classifier architecture.
Key Results
- On MNIST, over 70% of mislabeled images were detected, leading to a 5% increase in classification accuracy. Similar improvements were observed on Fashion-MNIST and CIFAR-10, with detection rates of 68% and 66%, respectively. The PSNR of reconstructed images correlated positively with detection success, confirming the effectiveness of latent space features. Compared to baseline methods like KNN and PCA, the autoencoder-based approach achieved superior noise removal, especially under high noise conditions, demonstrating robustness and scalability.
- The method's ability to identify and remove mislabeled data improves model robustness, reduces bias, and enhances generalization. It outperforms traditional noise filtering techniques by not relying on label consistency, making it suitable for real-world noisy datasets.
Significance
This work addresses a critical challenge in supervised learning—label noise—by proposing a simple yet effective pre-processing step that enhances data quality without altering existing models. It offers a scalable solution applicable across domains such as medical imaging, autonomous driving, and large-scale data annotation, where label errors are prevalent. By enabling unsupervised detection of mislabeled samples, it reduces reliance on costly manual verification, accelerates data cleaning, and ultimately improves model reliability and performance in real-world scenarios.
Technical Contribution
The key technical innovation lies in integrating convolutional autoencoders with adaptive density clustering to detect label errors in an unsupervised manner. The autoencoder compresses images into a meaningful latent space, preserving essential features while discarding noise. Rahmah’s algorithm dynamically determines optimal DBSCAN parameters for each class, ensuring accurate outlier detection. Unlike existing methods that depend on label consistency or model-based noise correction, this approach operates solely on feature distributions, making it broadly applicable and computationally efficient. Extensive experiments validate its effectiveness, setting a new standard for label noise mitigation.
Novelty
This study is the first to leverage the latent space representations from convolutional autoencoders for label noise detection, combined with adaptive density clustering. Unlike prior work relying on loss functions or model modifications, it offers a purely unsupervised, pre-processing solution. Its ability to detect mislabeled samples without additional supervision or complex architecture modifications marks a significant advancement in data cleaning techniques for supervised learning.
Limitations
- The approach assumes that samples with similar features share the same label; in highly imbalanced or complex datasets, this assumption may break down, reducing detection accuracy.
- Selection of latent space dimensionality and clustering parameters remains sensitive; improper choices can lead to false positives or missed errors.
- In scenarios with extremely high noise levels (>80%), detection performance diminishes, requiring further refinement.
Future Work
Future research could explore multi-modal data integration, active learning to iteratively refine label correctness, and deep generative models to enhance feature representation. Extending the framework to handle class imbalance, multi-label scenarios, and real-time data streams will broaden its applicability. Additionally, combining this pre-processing with robust training algorithms could further improve model resilience against noisy labels.
AI Executive Summary
Supervised learning models are highly dependent on the quality of labeled datasets. In real-world applications, labels often contain errors due to human mistakes or automated annotation processes, which can significantly degrade model performance. Traditional approaches to mitigate label noise involve modifying loss functions or adjusting model architectures, but these methods often require complex modifications and may not generalize well across different datasets.
This paper proposes a novel pre-processing framework that leverages the power of convolutional autoencoders combined with density-based clustering (DBSCAN) to identify and remove mislabeled images before training. The process begins by encoding images into a low-dimensional latent space where features are preserved. Using Rahmah’s algorithm, optimal clustering parameters are adaptively determined for each class, ensuring robust outlier detection. Outliers—samples with features deviating from the main cluster—are flagged as potential mislabeled data and eliminated from the dataset.
Extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate that this approach detects over 67% of mislabeled images, significantly improving the quality of training data. The detection accuracy surpasses traditional methods like KNN and PCA, especially under high noise conditions. The reconstructed images' PSNR correlates with detection success, validating the effectiveness of the feature representation.
By removing noisy labels, the method enhances the robustness and generalization of subsequent classifiers, leading to higher accuracy and stability. Its simplicity—requiring no changes to existing models—makes it highly practical for large-scale deployment in industry and research. Future work will focus on extending this framework to handle class imbalance, multi-label data, and real-time noisy data streams, further broadening its impact in data-centric AI development.
Deep Dive
Abstract
Supervised learning is based on the assumption that the ground truth in the training data is accurate. However, this may not be guaranteed in real-world settings. Inaccurate training data will result in some unexpected predictions. In image classification, incorrect labels may cause the classification model to be inaccurate as well. In this paper, I am going to apply unsupervised techniques to the training data before training the classification network. A convolutional autoencoder is applied to encode and reconstruct images. The encoder will project the image data on to latent space. In the latent space, image features are preserved in a lower dimension. The assumption is that data samples with similar features are likely to have the same label. Noised samples can be classified in the latent space by the Density-Base Scan (DBSCAN) clustering algorithm. These incorrectly labeled data are visualized as outliers in the latent space. Therefore, the outliers identified by the DBSCAN algorithm can be classified as incorrectly labeled samples. After the outliers are detected, all the outliers are treated as mislabeled data samples and removed from the dataset. Thus the training data can be directly used in training the supervised learning network. The algorithm can detect and remove above 67\% of mislabeled data in the experimental dataset.