Grigoris

Grigoris

About me

Welcome to my site. My name is Grigoris and I am an Assistant Professor in University of Wisconsin-Madison.

My research focuses on reliable machine learning, focusing on the development of robust models that perform well under noise and out-of-distribution data. Concretely,

  • Architecture design: I have worked extensively on polynomial networks (PNs) that capture high-degree interactions between inputs. My short-term goals are to understand the inductive bias and properties of existing architectures through empirical and theoretical studies. I am interested in the complete theoretical understanding of (neural/polynomial) networks, including their expressivity, trainability, generalization properties, and inductive biases. Our recent work has provided the first characterization of the generalization of this class of functions, or the spectral bias of high-degree polynomials, highlighting how PNs can learn higher frequency functions faster than regular feed-forward networks.
  • Trustworthy models: My goal is to understand the enhance the performance of existing networks, particularly with respect to their extrapolation abilities, and their robustness to malicious (adversarial) attacks. I am interested in both discriminative and generative models, including Large Language Models. In our recent work, we have studied adversarial attacks and defenses in the text domain, where there are exciting questions ahead. Our long-term goal is to develop models that are robust, fair, and capable of generalizing well to unseen combinations with strong extrapolation abilities.

News

Funding Acknowledgement

I would like to acknowledge the funding of the following organizations who have generously supported various events or projects in the past. I am very thankful for their support:
  • 2024: Google and OpenAI: grants on trustworthy Large Language Models (LLMs).
  • 2024: ELISE Fellows Mobility Program: travel grant for short-term visit of an ELLIS lab.