- • Collaborated with researchers at Oxford University to propose and experimentally validate a novel methodology for training Reinforcement Learning (RL) agents that do not prefer to prevent their own shutdown when given the opportunity, contributing to the field of AI safety
- • Developed a novel training scheme and reward function to promote agent neutrality without compromising performance. Designed new behavioral metrics to evaluate agent alignment and usefulness
Training Shutdownable Agents via Stochastic Choice
Published: Proceedings of the Technical AI Safety Conference 2025
- • Formulated and executed on a novel research agenda, introducing a methodology for discovering algebras of symmetries present in Neural Networks, leading to the publication of five papers on the topic in 2023
- • Validated the methodology by autonomously rediscovering non-trivial properties of classical Lie groups present in a toy model, laying the groundwork for future research on interpretability using symmetries
- • Extended the methodology to discover non-trivial symmetries of a deep neural network
- • Explored methods to improve robustness by regularizing models using non-trivial symmetries
Identifying the Group-Theoretic Structure of Machine-Learned Symmetries
Published: Physics Letters B
Accelerated Discovery of Machine-Learned Symmetries: Deriving the Exceptional Lie Groups G₂, F₄ and E₆
Published: Physics Letters B
Discovering Sparse Representations of Lie Groups with Machine Learning
Published: Physics Letters B
Oracle Preserving Latent Flows
Published: Symmetry
Deep Learning Symmetries and Their Lie Groups, Algebras, and Subalgebras from First Principles
Published: Machine Learning: Science and Technology
- • Won 1st place in the 2022 NeurIPS Ariel Data Challenge
- • Predicted atmospheric properties of distant exoplanets from their spectra using physics informed AI, which could aid in detecting signatures of biological life on alien worlds. Featured in AAS Nova
- • Developed framework which reduces the dimensionality of the problem by exploiting fundamental physics principles, utilizing inductive biases from domain expertise and improving the interpretability of the analysis
- • Utilized unsupervised learning to discover unknown features on the manifold of transit spectra
Lessons Learned from Ariel Data Challenge 2022: Inferring Physical Properties of Exoplanets From Next-Generation Telescopes
Published: Proceedings of Machine Learning Research
Transverse Vector Decomposition Method for Analytical Inversion of Exoplanet Transit Spectra
Published: The Astrophysical Journal
Unsupervised Machine Learning for Exploratory Data Analysis of Exoplanet Transmission Spectra
Published: The Planetary Science Journal
Analytical Modeling of Exoplanet Spectroscopy with Dimensional Analysis and Symbolic Regression
Published: The Astrophysical Journal
- • Derived information theoretic bounds on the statistical uncertainty of parameter estimations using GAN-generated datasets, to be used in the analysis of LHC data, and implemented GAN architecture
Uncertainties associated with GAN-generated datasets in high energy physics
Published: SciPost Physics