CRII: RI: Uncertainty Estimation and Robustness in Hierarchical Classification

Hierarchical classification is the task of categorizing items belonging to a structured hierarchy. Examples of this task include tracking of animals for monitoring variability in species and pests in precision agriculture. Reliable automation of hierarchical classification has the potential to accelerate sustainability efforts and boost agricultural productivity. Machine learning is increasingly being used as an automation tool in ecology and agriculture. However, to fully realize the potential benefits of machine learning in hierarchical classification, principled methodologies are needed to ensure the trustworthiness and robustness of the models used in these applications. This research will develop methodologies for machine learning models to accurately report their uncertainty about predictions and to defer to human experts in a principled, resource-aware manner. Data scarcity and low-quality labels often hinder the effectiveness of machine learning systems. To address these challenges, this research will also develop theory and methodologies to overcome the constraints of scarce and low-quality data, both of which are common in hierarchical classification tasks in practice. The project will develop isotonic regression methods tailored for uncertainty quantification in hierarchical classification. Moreover, it will design efficient algorithms for learning-to-defer in hierarchical classification. It will also develop hierarchical few-shot learning techniques to address data scarcity and robust algorithms for handling hierarchical label noise. Collectively, these efforts aim to establish principled methodologies that ensure the trustworthiness and robustness of hierarchical classification systems, enabling their effective use in mission-critical applications constrained by scarce, low-quality data. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.