Convergent Learning for Class Imbalance: A Unified Approach to Long-Tail Recognition in Image Classification

Soliman Hanna, Bishoy

Student Work

Convergent Learning for Class Imbalance: A Unified Approach to Long-Tail Recognition in Image Classification

Public Deposited

Addressing the complex challenge of long-tail recognition (LTR) in image classification, this study introduces an innovative integrated approach that meticulously combines cross-entropy, contrastive learning, and imbalanced class loss functions. The prevalence of class imbalances within datasets significantly hampers the effectiveness of conventional machine learning models by skewing performance toward majority classes and neglecting the minority ones. To combat this issue, our approach aims to harmonize the learning process across all classes, ensuring equitable representation and enhancing feature separability. Our methodology is rooted in a multifaceted strategy where metric learning combined with crossentropy fosters an equitable learning environment across diverse classes, ensuring no class is overshadowed regardless of its frequency within the dataset. This composite loss function is designed to bolster model robustness and generalization capabilities, addressing the inherent challenges of the LTR problem comprehensively. The validation of our proposed solution was conducted through rigorous experimentation on modified CIFAR-10/100 datasets and a bespoke custom dataset, showcasing our approach’s adaptability and effectiveness across varying levels of class imbalance. Utilizing ResNet models of differing depths (ResNet18, ResNet34, and ResNet50) and experimenting with a range of loss functions—including focal loss and supervised contrastive loss—allowed us to assess our methodology’s performance in a broad spectrum of scenarios. This experimental setup not only facilitated a deep dive into the comparative analysis of model behaviors but also enabled the identification of optimal configurations for tackling LTR challenges. Our findings illustrate significant improvements in model performance, particularly in environments characterized by pronounced class imbalances. By employing our integrated approach, we were able to set new benchmarks for image classification models, demonstrating superior performance in handling longtailed distributions. The study’s implications extend beyond the immediate advancements in LTR; it lays a foundational framework for future research in machine learning, emphasizing the importance of a balanced and nuanced approach to model training and development. Furthermore, our study elucidates the critical role of selecting appropriate loss functions and data augmentation strategies, tailored to the unique characteristics of each dataset. This insight is instrumental in advancing the field of machine learning, guiding practitioners in the development of more robust, equitable, and effective models. Through a comprehensive exploration of the challenges and solutions associated with LTR, this research contributes to a deeper understanding of class imbalance issues, paving the way for innovative approaches in the domain of imbalanced learning.

This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator