A task-oriented few-shot spectroscopy benchmark: Empirical evaluation of meta-learning approaches

Abstract

Near- and mid-infrared (NIR/MIR) spectroscopy, crucial for rapid, noninvasive chemical analysis in industrial applications, is increasingly leveraged by machine learning techniques to extract meaningful information from complex spectral data. Nevertheless, the constraints imposed by limited sample sizes and the occurrence of domain shifts present substantial obstacles to the effectiveness of both deep learning and transfer learning methodologies. In this work, we introduce a unified, task-oriented benchmark that consolidates nine open-access spectral datasets, providing a consistent preprocessing pipeline and clearly defined few-shot tasks. The benchmark is segmented into multiple tests exhibiting varying degrees of heterogeneity, thereby facilitating the evaluation of diverse methodologies across a range of contexts characterized by varying levels of complexity. To the best of our knowledge, meta-learning has not previously been applied in the context of spectrometric analysis. In this study, we confirm that it offers superior performance to traditional transfer learning approaches. A systematic comparison of transfer learning and meta-learning techniques reveals that meta-learning consistently outperforms transfer learning—achieving lower errors in both heterogeneous and homogeneous tasks, consistent with our latent-space analysis showing a more structured cross-task embedding geometry that helps explain these gains. Furthermore, we empirically demonstrate that this benchmark can serve as a starting point for pre-training models on new industrial problems, accelerating adaptation and improving final performance using meta-learning. We complement our analysis with established, model-agnostic interpretability tools, enhancing transparency and industrial relevance in regulated domains.