Adaptive loss optimization for enhanced learning performance: application to image-based rock classification

Soroor Salavati, Pedro Ribeiro Mendes Júnior, Anderson Rocha, Alexandre Ferreira

Abstract: In the geoscience domain, mainly within the oil and gas industry, getting the correct category of rock samples is crucial. Machine learning models deployed for rock classification often use the categorical cross-entropy loss function. This loss function may struggle when the rocks are either very similar or have too much intraclass variability. Furthermore, categorical cross-entropy loss may ignore subtle but significant differences between classes. This results in the ignoring of hard-to-classify samples and fails to prioritize learning from these challenging patterns. This can lead to models biased toward more common classes or mislabeling of underrepresented rock types, especially when dealing with noisy or inconsistent data. To bridge those gaps, we propose a new hybrid loss function. It combines the traditional categorical cross-entropy loss function with Online Hard Example Mining (OHEM), a method originally formulated for object detection tasks focusing on hard-to-classify samples. We designed this function to be adjustable, making it adaptable to various challenges inherent to the rock classification. We evaluated this technique on a diverse, highly heterogeneous, and challenging dataset provided by Shell Brazil. The available techniques were tested and encountered problems with challenging classes. However, our new technique improved classification accuracy for both challenging and easier samples. In addition to rock classification, this technique may serve as a blueprint for addressing complicated classification problems in other fields.