MoMo Strategy : Learn More from More Mistakes

Sophia, Chulif and Lee, Sue Han and Chang, Yang Loong and Mark Tsun, Tee Kit and Chai, Kok Chin and Then, Yi Lung (2023) MoMo Strategy : Learn More from More Mistakes. In: 2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 31 October 2023 - 03 November 2023, Taiwan.

PDF MoMo_Strategy_Learn_More_from_More_Mistakes.pdf Download (24MB)

Abstract

Training accurate convolutional neural networks (CNNs) is essential for achieving high-performance machine learning models. However, limited training data pose a challenge, reducing model accuracy. This research investigates the selection and utilization of misclassified training samples to enhance the accuracy of CNNs where the dataset is long-tail distributed. Unlike classical resampling methods involving oversampling of tail classes and undersampling of head classes, we propose an approach that allocates more misclassified training samples into the training process to learn more (namely, MoMo strategy), with ratios of 50:50 and 70:30 for the wrongly predicted and correctly predicted samples, respectively. Additionally, we propose incorporating a balanced sample selection method, whereby the maximum training sample per class in an epoch is assigned to address the long-tail dataset problem. Our experimental results on a subset of the current largest plant dataset, PlantCLEF 2023, demonstrate an increase of 1%-2% in overall validation accuracy and a 2%-5% increase in tail class identification. By selectively focusing on more misclassified samples in training, at the same time, integrating a balanced sample selection achieves a significant boost in accuracy compared to traditional training methods. These findings emphasize the significance of adding more misclassified samples into training, encouraging researchers to rethink the sampling strategies before implementing more complex and robust network architectures and modules.

Actions (For repository members only: login required)

View Item