Katherine Atwell

Katherine Atwell

Final-year Computer Science PhD candidate @ Northeastern University

News

  • 08.23.25: Our work, BASIL: Bayesian Assessment of Sycophancy in LLMs, was published at ArXiv
  • 07.31.25: I am co-organizing the fourth NLP for Positive Impact Workshop, co-located at ACL 2025!
  • 07.27.25: Our work, Measuring Bias and Agreement in Large Language Model Presupposition Judgments, was published at ACL Findings 2024
  • 07.27.25: Our work, Contextual ASR Error Handling with LLMs Augmentation for Goal-Oriented Conversational AI, was published at COLING Industry Track 2025
  • 06.25.25: My advisor, Malihe Alikhani, spoke with a Wall Street Journal reporter about our lab's work on AI sycophancy; see the article here!
  • 11.12.24: Our work, Studying and Mitigating Biases in Sign Language Understanding Models, was published at EMNLP 2024
  • 06.26.24: Our work, The Importance of Including Signed Languages in Natural Language Processing, was published at
  • 06.16.24: Our work, Generating Signed Language Instructions in Large-Scale Dialogue Systems, was published at NAACL 2024
See more

PAC-bayesian domain adaptation bounds for multiclass learners

Published in In the proceedings of Uncertainty in Artificial Intelligence, 2022

Recommended citation: Anthony Sicilia, Katherine Atwell, Malihe Alikhani, Seong Hwang, "PAC-bayesian domain adaptation bounds for multiclass learners." In the proceedings of Uncertainty in Artificial Intelligence, 2022.

Multiclass neural networks are a common tool in modern unsupervised domain adaptation, yet an appropriate theoretical description for their non-uniform sample complexity is lacking in the adaptation literature. To fill this gap, we propose the first PAC-Bayesian adaptation bounds for multiclass learners. We facilitate practical use of our bounds by also proposing the first approximation techniques for the multiclass distribution divergences we consider. For divergences dependent on a Gibbs predictor, we propose additional PAC-Bayesian adaptation bounds which remove the need for inefficient Monte-Carlo estimation. Empirically, we test the efficacy of our proposed approximation techniques as well as some novel design-concepts which we include in our bounds. Finally, we apply our bounds to analyze a common adaptation algorithm that uses neural networks.