Image of Dirichlet Classifiers charts
The Department of Mathematical and Statistical Sciences is proud to announce that PhD student Courtney Franzen has had her paper accepted for an oral presentation at the prestigious IEEE International Conference on Machine Learning and Applications (ICMLA). This is a remarkable achievement, as only 25% of submissions were accepted for oral presentation in this highly competitive track.
The paper titled “Out-of-the-Box Uncertainty: Reducing Confidence Errors with Dirichlet Classifiers”, was co-authored with her advisor, Dr. Farhad Pourkamali-Anaraki.
Overconfident misclassifications remain a critical barrier to deploying deep classifiers in risk-sensitive applications. Franzen’s research presents a targeted empirical comparison between two classification models: one trained with cross-entropy loss and temperature scaling (CE classifiers) and another trained using a Dirichlet-based surrogate loss with an annealed Kullback–Leibler (KL) divergence penalty (Dirichlet classifiers). To enable stable Dirichlet training in high-class-count regimes, the study introduces a dataset-agnostic KL annealing schedule that mitigates training collapse—an essential step for tasks involving fine-grained recognition, multi-label categorization, or large taxonomies where class counts are high and reliable confidence estimation is critical.
Evaluated across CIFAR-5 to CIFAR-100, the results show that Dirichlet classifiers consistently reduce high-confidence misclassifications by 2–5x compared to CE baselines, while maintaining similar accuracy.
This accomplishment highlights not only Courtney’s dedication and innovative research but also the strong mentorship of Dr. Pourkamali-Anaraki. Together, their work represents a significant contribution to advancing machine learning methods in risk-sensitive applications.
We congratulate Courtney Franzen on this outstanding achievement and extend our appreciation to Dr. Farhad Pourkamali-Anaraki for his guidance and support.