Parameter-Free Loss for Class-Imbalanced Deep Learning in Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models often struggle with class-imbalanced datasets, necessitating complex loss functions.
• Existing loss functions require extensive hyperparameter tuning, hindering training efficiency and user accessibility.

Purpose of the Study:

• To introduce a parameter-free loss (PF-loss) function for binary and multiclass imbalanced deep learning image classification.
• To address the limitations of current loss functions by eliminating the need for hyperparameter optimization.

Main Methods:

• A novel parameter-free loss (PF-loss) function was developed for deep learning image classification.
• The PF-loss function dynamically focuses on minority classes without requiring hyperparameter adjustments.
• The method adapts to changing data distributions within mini-batches during training.

Main Results:

• PF-loss significantly reduces training time, achieving up to a 1/148 reduction compared to state-of-the-art methods.
• The proposed loss function achieves comparable or higher accuracy, measured by G-mean and AUC metrics.
• Performance improvements are particularly notable in highly skewed datasets with imbalance ratios up to 200.

Conclusions:

• The parameter-free loss function offers a more efficient and practical solution for training deep learning models on imbalanced image classification tasks.
• PF-loss enhances model performance by dynamically adapting to data distribution shifts without hyperparameter tuning.
• This approach democratizes the use of advanced deep learning techniques for non-expert users dealing with imbalanced data.