arXiv:2509.23964v2 Announce Type: replace-cross 
Abstract: Label noise in datasets could significantly damage the performance and robustness of deep neural networks (DNNs) trained on these datasets. As the size of modern DNNs grows, there is a growing demand for automated tools for detecting such errors. In this paper, we propose post-hoc, model-agnostic noise detection and rectification methods utilizing the penultimate feature from a DNN. Our idea is based on the observation that the similarity between the penultimate feature of a mislabeled data point and its true class data points is higher than that for data points from other classes, making the probability of label occurrence within a tight, similar cluster informative for detecting and rectifying errors. Through theoretical and empirical analyses, we demonstrate that our approach achieves high detection performance across diverse, realistic noise scenarios and can automatically rectify these errors to improve dataset quality. Our implementation is available at https://anonymous.4open.science/r/noise-detection-and-rectification-AD8E.

تسلط دراسة حديثة الضوء على نهج واعد لمعالجة ضوضاء التسميات في مجموعات البيانات، والتي يمكن أن تعيق أداء الشبكات العصبية العميقة. من خلال تقديم طرق غير مرتبطة بالنموذج لاكتشاف وتصحيح هذه الأخطاء، تتناول الأبحاث حاجة ملحة في مجال التعلم الآلي. لا تعزز هذه الخطوة موثوقية الشبكات العصبية العميقة فحسب، بل تمهد أيضًا الطريق لتطبيقات أكثر قوة في مجالات متنوعة، مما يجعلها مساهمة مهمة في التطور المستمر لتكنولوجيا الذكاء الاصطناعي.

Un estudio reciente destaca un enfoque prometedor para abordar el ruido en las etiquetas de los conjuntos de datos, que puede obstaculizar el rendimiento de las redes neuronales profundas. Al introducir métodos agnósticos al modelo para detectar y corregir estos errores, la investigación aborda una necesidad crítica en el campo del aprendizaje automático. Este avance no solo mejora la fiabilidad de las DNN, sino que también allana el camino para aplicaciones más robustas en diversos dominios, convirtiéndose en una contribución significativa a la evolución continua de la tecnología de IA.

Une étude récente met en avant une approche prometteuse pour lutter contre le bruit des étiquettes dans les ensembles de données, qui peut nuire à la performance des réseaux de neurones profonds. En introduisant des méthodes agnostiques au modèle pour détecter et corriger ces erreurs, la recherche répond à un besoin crucial dans le domaine de l'apprentissage automatique. Cette avancée améliore non seulement la fiabilité des DNN, mais ouvre également la voie à des applications plus robustes dans divers domaines, ce qui en fait une contribution significative à l'évolution continue de la technologie IA.

A recent study highlights a promising approach to tackle label noise in datasets, which can hinder the performance of deep neural networks. By introducing model-agnostic methods for detecting and correcting these errors, the research addresses a critical need in the field of machine learning. This advancement not only enhances the reliability of DNNs but also paves the way for more robust applications in various domains, making it a significant contribution to the ongoing evolution of AI technology.

Detecting and Rectifying Noisy Labels: A Similarity-based Approach

arXiv:2511.17103v1 Announce Type: new 
Abstract: Visual emotion recognition (VER) is a longstanding field that has garnered increasing attention with the advancement of deep neural networks. Although recent studies have achieved notable improvements by leveraging the knowledge embedded within pre-trained visual models, the lack of direct association between factual-level features and emotional categories, called the "affective gap", limits the applicability of pre-training knowledge for VER tasks. On the contrary, the explicit emotional expression and high information density in textual modality eliminate the "affective gap". Therefore, we propose borrowing the knowledge from the pre-trained textual model to enhance the emotional perception of pre-trained visual models. We focus on the factual and emotional connections between images and texts in noisy social media data, and propose Partitioned Adaptive Contrastive Learning (PACL) to leverage these connections. Specifically, we manage to separate different types of samples and devise distinct contrastive learning strategies for each type. By dynamically constructing negative and positive pairs, we fully exploit the potential of noisy samples. Through comprehensive experiments, we demonstrate that bridging the "affective gap" significantly improves the performance of various pre-trained visual models in downstream emotion-related tasks. Our code is released on https://github.com/wdqqdw/PACL.

تقدم دراسة حديثة نهجًا مبتكرًا للتعرف على المشاعر البصرية (VER) من خلال اقتراح استخدام التعلم التبايني التكيفي المقسم (PACL) لسد 'الفجوة العاطفية' بين الأنماط البصرية والنصية. تستفيد هذه الطريقة من المعرفة المستمدة من النماذج النصية المدربة مسبقًا لتعزيز قدرات الإدراك العاطفي للنماذج البصرية، خاصة في سياقات وسائل التواصل الاجتماعي المليئة بالضوضاء.

Un estudio reciente presenta un enfoque novedoso para el reconocimiento de emociones visuales (VER) al proponer el uso del Aprendizaje Contrastivo Adaptativo Partitionado (PACL) para cerrar la 'brecha afectiva' entre las modalidades visual y textual. Este método aprovecha el conocimiento de modelos textuales preentrenados para mejorar las capacidades de percepción emocional de los modelos visuales, especialmente en contextos de redes sociales ruidosas.

Une étude récente présente une approche novatrice pour la reconnaissance des émotions visuelles (VER) en proposant l'utilisation de l'apprentissage contrastif adaptatif partitionné (PACL) pour combler le 'écart affectif' entre les modalités visuelles et textuelles. Cette méthode exploite les connaissances des modèles textuels pré-entraînés pour améliorer les capacités de perception émotionnelle des modèles visuels, en particulier dans des contextes de médias sociaux bruyants.

A recent study introduces a novel approach to visual emotion recognition (VER) by proposing the use of Partitioned Adaptive Contrastive Learning (PACL) to bridge the 'affective gap' between visual and textual modalities. This method leverages knowledge from pre-trained textual models to enhance the emotional perception capabilities of visual models, particularly in noisy social media contexts.

Bridging Visual Affective Gap: Borrowing Textual Knowledge by Learning from Noisy Image-Text Pairs

arXiv:2509.05826v2 Announce Type: replace 
Abstract: Conformal prediction is a model-agnostic approach to generating prediction sets that cover the true class with a high probability. Although its prediction set size is expected to capture aleatoric uncertainty, there is a lack of evidence regarding its effectiveness. The literature presents that prediction set size can upper-bound aleatoric uncertainty or that prediction sets are larger for difficult instances and smaller for easy ones, but a validation of this attribute of conformal predictors is missing. This work investigates how effectively conformal predictors quantify aleatoric uncertainty, specifically the inherent ambiguity in datasets caused by overlapping classes. We perform this by measuring the correlation between prediction set sizes and the number of distinct labels assigned by human annotators per instance. We further assess the similarity between prediction sets and human-provided annotations. We use three conformal prediction approaches to generate prediction sets for eight deep learning models trained on four datasets. The datasets contain annotations from multiple human annotators (ranging from five to fifty participants) per instance, enabling the identification of class overlap. We show that the vast majority of the conformal prediction outputs show a very weak to weak correlation with human annotations, with only a few showing moderate correlation. These findings underscore the necessity of critically reassessing the prediction sets generated using conformal predictors. While they can provide a higher coverage of the true classes, their capability in capturing aleatoric uncertainty and generating sets that align with human annotations remains limited.

يتم تقييم التنبؤ المتوافق، وهو نهج مستقل عن النموذج، لفعاليته في التقاط عدم اليقين العشوائي، والذي يشير إلى الغموض الكامن في مجموعات البيانات بسبب تداخل الفئات. تركز هذه الدراسة على العلاقة بين أحجام مجموعات التنبؤ وعدد التسميات المميزة التي يخصصها المعلقون البشر لكل حالة، بهدف التحقق من الأداء المتوقع لمتنبئي التنبؤ المتوافق في البيئات غير المؤكدة.

La predicción conformal, un enfoque independiente del modelo, se está evaluando por su efectividad en la captura de la incertidumbre aleatoria, que se refiere a la ambigüedad inherente en los conjuntos de datos debido a clases superpuestas. Esta investigación se centra en la correlación entre los tamaños de los conjuntos de predicción y las etiquetas distintas asignadas por anotadores humanos, con el objetivo de validar el rendimiento esperado de los predictores conformales en entornos inciertos.

La prédiction conforme, une approche indépendante du modèle, est évaluée pour son efficacité à capturer l'incertitude aléatoire, qui fait référence à l'ambiguïté inhérente dans les ensembles de données en raison de classes qui se chevauchent. Cette enquête se concentre sur la corrélation entre les tailles des ensembles de prédiction et les étiquettes distinctes attribuées par des annotateurs humains, visant à valider la performance attendue des prédicteurs conformes dans des environnements incertains.

Conformal prediction, a model-agnostic approach, is being evaluated for its effectiveness in capturing aleatoric uncertainty, which refers to the inherent ambiguity in datasets due to overlapping classes. This investigation focuses on the correlation between prediction set sizes and the distinct labels assigned by human annotators, aiming to validate the expected performance of conformal predictors in uncertain environments.

Detecting and Rectifying Noisy Labels: A Similarity-based Approach

Was this article worth reading? Share it

All Voice Lab

Meteoria

ZeroGPT.org