Σεμινάριο: "Learning to Align: Addressing Frequency Distribution Shifts"

ΚΥΚΛΟΣ ΣΕΜΙΝΑΡΙΩΝ ΣΤΑΤΙΣΤΙΚΗΣ 2025-2026

Ομιλητής: John Pavlopoulos, Assistant Professor, Department of Informatics, Athens University of Economics and Business and Archimedes, Athena Research Center, Greece

Αμφιθέατρο Τροίας

ΠΕΡΙΛΗΨΗ

This talk focuses on a novel loss function that incorporates the Wasserstein distance between frequency distribution of the predicted text and a target distribution empirically derived from training data. By penalizing divergence from expected distributions, our approach enhances both accuracy and robustness under temporal and contextual intra-dataset shifts. To showcase the advantage of this approach, we use handwritten text recognition (HTR), where visual input is converted into machine-readable text. This is a challenging domain due to the evolving and context-dependent nature of handwriting. Character sets change over time, and character frequency distributions shift across historical periods or regions, often causing models trained on broad, heterogeneous corpora to underperform on specific subsets. As we show, character distribution alignment can improve existing models at inference time without requiring retraining by integrating it as a scoring function in a guided decoding scheme. Experimental results across multiple datasets and architectures confirm the effectiveness of our method in boosting generalization and performance. Our code is publicly available: https://github.com/pkaliosis/fada.

Joint work with Panagiotis Kalliosis (PhD student, Stony Brook University, New York, USA).