What question did this study set out to answer?

Design an effective deep neural network architecture that leverages a minimal amount of clean meta-data to achieve robust image classification despite high proportions of noisy training labels.

February 16, 2025Applied SciencesOpen Access

Meta-Data-Guided Robust Deep Neural Network Classification with Noisy Label

View Full Paper

Ask AI

Bookmark

Authors

JLJie LüUniversity of Technology Sydney YWYufeng WangJining Medical University ASAiju ShiNanjing University of Posts and Telecommunications

Discussion

Member takes

Implication

Algorithm evaluation demonstrates improved classification accuracy across noisy datasets using an asymmetric dual-branch framework, highlighting the utility of targeted meta-data guidance.

Key Points

Design an effective deep neural network architecture that leverages a minimal amount of clean meta-data to achieve robust image classification despite high proportions of noisy training labels.
Constructed a cross dual-branch network (CoNet-MS) where each branch applies a small-loss criterion to select pseudo-clean samples to cross-train the companion branch.
Integrated a single-sided meta-guided module to dynamically weight observed versus model-predicted labels in the loss function, generating branch asymmetry to mitigate confirmation bias.
Benchmarked classification performance across multiple datasets subjected to synthetic and real-world label noise against baseline and state-of-the-art learning-with-noisy-label methods.
CoNet-MS outperformed existing learning-with-noisy-label methods, including state-of-the-art meta-data algorithms, across various noise types and label noise ratios.
The asymmetric cross-branch interaction successfully reduced error accumulation from mislabeled samples and prevented models from reinforcing self-training errors.

Cite This Study

Lü et al. (2025) studied this question.

synapsesocial.com/papers/6a155c1679ff98d0de4e81ab https://doi.org/10.3390/app15042080

View Full Paper

Ask AI

Bookmark

Implication

Algorithm evaluation demonstrates improved classification accuracy across noisy datasets using an asymmetric dual-branch framework, highlighting the utility of targeted meta-data guidance.

Meta-Data-Guided Robust Deep Neural Network Classification with Noisy Label

Discussion

Member takes

Implication

Key Points

Cite This Study

Meta-Data-Guided Robust Deep Neural Network Classification with Noisy Label

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider