Alizée Pace

Alizée Pace

PhD Student in Machine Learning

ETH Zürich & MPI-IS

About Me

Hi! I’m Alizée, a PhD student in Machine Learning and a fellow at the ETH AI Center in Zürich. I am lucky to work with Prof. Gunnar Rätsch and Prof. Bernhard Schölkopf. I am also part of the ELLIS PhD program. My main research goal is to develop ML solutions for decision support systems, with an interest for offline reinforcement learning, preference-based RL and real-world clinical applications.

I have been a research intern (2024) and student researcher (2023) at Google, where I developed new methods in Reinforcement Learning from Human Feedback (RLHF) to improve LLM model quality and alignment within the Gemini team.

Before my PhD, I worked on interpretable imitation learning with Prof. Mihaela van der Schaar at the University of Cambridge. My professional experience also includes medical device development and software engineering at CERN. I studied Physics, Materials Science and Machine Learning at Cambridge, where I consistently ranked first in my year.

Interests
  • Offline RL & Imitation
  • RLHF
  • Causal Inference
  • Clinical Time-Series
Education
  • PhD in Machine Learning, started 2021

    ETH Zürich

  • MPhil in Machine Learning and Machine Intelligence, 2021

    University of Cambridge

  • BA MSci in Materials Science, 2020

    University of Cambridge

Recent News

All news»

[July 26, 2024] I am co-organising a workshop on RL theory & practice at ICML 2024.

[June 26, 2024] My paper on preference elicitation for offline RL is out on arXiv.

[June 3, 2024] I am back at Google as a Research Intern, working on RLHF and reward modelling for Gemini.

[Jan 22, 2024] My paper on synthetic preference generation, based on work during my internship at Google, is out on arXiv.

[Jan 16, 2024] My paper on delphic offline RL is accepted to ICLR 2024.

Professional Experience

 
 
 
 
 
Research Intern
Jun 2024 – Sep 2024 Zürich, Switzerland
Developing new methods to improve large language model (LLM) quality within the Gemini team. Research focused on reinforcement learning from human feedback (RLHF) and reward modelling.
 
 
 
 
 
Student Researcher
May 2023 – Nov 2023 Zürich, Switzerland
Developed new methods to improve large language model (LLM) quality within the Gemini team. Research focused on reinforcement learning from human feedback (RLHF) and reward modelling. My contributions were integrated within the product and resulted in a patent application and a publication.
 
 
 
 
 
Doctoral Research Fellow
Sep 2021 – Present Zürich, Switzerland
My main research goal is to develop ML solutions for decision support and treatment recommendation systems. General areas of interest include offline reinforcement learning, representation learning, causal inference and time-series modelling.
 
 
 
 
 
Research Assistant - Machine Learning
Mar 2021 – Sep 2021 Cambridge, UK
Research project on interpretable imitation learning for clinical decision support. Our goal was to describe and understand treatment or diagnostic policies through novel decision tree models, and capture how decision-making behaviour varies over time with patient information. Publication accepted as Spotlight for ICLR 2022 (5% acceptance rate).
 
 
 
 
 
Research Assistant - Bioelectronics
Sep 2019 – Mar 2020 Cambridge, UK
Research project on printed biocompatible force sensors for orthopaedic implants. Resulted in two publications (1 and 2).
 
 
 
 
 
R&D Engineering Intern - Medical Device Data Analysis
Jul 2019 – Sep 2019 Barcelona, Spain
Development of a new thrombectomy device which restores blood flow in stroke patients. Resulted in three patent applications (1, 2 and 3)
 
 
 
 
 
Software Development Intern
Aug 2016 – Aug 2016 Geneva, Switzerland
Development of Invenio user interface and of demos for a reproducible analysis platform.

Publications

See my Google Scholar profile for a full list of publications.

(2024). Delphic Offline Reinforcement Learning under Nonidentifiable Hidden Confounding. ICLR 2024.

Cite Source Document

(2023). On the Importance of Step-wise Embeddings for Heterogeneous Clinical Time-Series. In ML4H 2023 (PMLR).

Cite Source Document

(2023). Temporal Label Smoothing for Early Event Prediction. In ICML 2023.

Cite Source Document

(2023). Clinical Trajectory Representations for Clustering. In ICLR TSRL4H 2023.

Cite Source Document

(2022). Conformable and robust force sensors to enable precision joint replacement surgery. In Materials & Design.

Cite Source Document

Contact