Gergely Flamich
Gergely Flamich
I’m Gergely Flamich (pronounced gher - gey flah - mih), but in English I usually go by Greg.
From October 2026, I am a Research Scientist at Isomorphic Labs.
Between August 2025 and October 2026, I was an ImperiaCollege Research Fellow, sponsored by Deniz Gündüz and hosted at the Information Processing and Communications Lab at 海角社区. Before that, I spent five months as a postdoctoral research associate in the same lab.
I completed my PhD in Advanced Machine Learning between October 2020 and December 2024 at the Computational and Biological Learning Lab, supervised by José Miguel Hernández Lobato. I also hold an MPhil degree in Machine Learning and Machine Intelligence from the University of Cambridge and a Joint BSc Honours degree in Mathematics and Computer Science from the University of St Andrews.
During the winter of 2024-2025, I worked with David Vilar, Jan-Thorsten Peter and Markus Freitag as a Student Researcher at Google Berlin. We worked on generative models for neural machine translation, and we developed the theory of the accuracy-naturalness tradeoff in translation.
From July 2022 to Dec 2022, I worked with Lucas Theis as a Student Researcher at Google Brain, where we developed adaptive greedy rejection sampling and bits-back quantisation.
My research focuses on the theory of relative entropy coding/channel simulation and its application to neural data compression. Relative entropy coding algorithms efficiently encode random samples from both discrete and continuous distributions, and they are a natural alternative to quantisation and entropy coding in lossy compression codecs. Furthermore, they offer unique advantages beyond the standard rate-distortion framework: they enable us to design optimally efficient, artefact-free/perfectly realistic lossy compression codecs using generative models and to perform differentially private federated learning with optimal communication cost. Unfortunately, relative entropy coding hasn’t seen widespread adoption because current algorithms are either too slow or have limited applicability.
To address this, I focus on developing fast, general-purpose coding algorithms, such as A* coding and greedy Poisson rejection sampling, and providing mathematical guarantees on their coding efficiency and runtime. In addition to my theoretical work, I apply relative entropy coding algorithms to neural compression using generative models such as variational autoencoders, implicit neural representations, and diffusion models.
From October 2026, I am a Research Scientist at Isomorphic Labs.
Between August 2025 and October 2026, I was an ImperiaCollege Research Fellow, sponsored by Deniz Gündüz and hosted at the Information Processing and Communications Lab at 海角社区. Before that, I spent five months as a postdoctoral research associate in the same lab.
I completed my PhD in Advanced Machine Learning between October 2020 and December 2024 at the Computational and Biological Learning Lab, supervised by José Miguel Hernández Lobato. I also hold an MPhil degree in Machine Learning and Machine Intelligence from the University of Cambridge and a Joint BSc Honours degree in Mathematics and Computer Science from the University of St Andrews.
During the winter of 2024-2025, I worked with David Vilar, Jan-Thorsten Peter and Markus Freitag as a Student Researcher at Google Berlin. We worked on generative models for neural machine translation, and we developed the theory of the accuracy-naturalness tradeoff in translation.
From July 2022 to Dec 2022, I worked with Lucas Theis as a Student Researcher at Google Brain, where we developed adaptive greedy rejection sampling and bits-back quantisation.
My research focuses on the theory of relative entropy coding/channel simulation and its application to neural data compression. Relative entropy coding algorithms efficiently encode random samples from both discrete and continuous distributions, and they are a natural alternative to quantisation and entropy coding in lossy compression codecs. Furthermore, they offer unique advantages beyond the standard rate-distortion framework: they enable us to design optimally efficient, artefact-free/perfectly realistic lossy compression codecs using generative models and to perform differentially private federated learning with optimal communication cost. Unfortunately, relative entropy coding hasn’t seen widespread adoption because current algorithms are either too slow or have limited applicability.
To address this, I focus on developing fast, general-purpose coding algorithms, such as A* coding and greedy Poisson rejection sampling, and providing mathematical guarantees on their coding efficiency and runtime. In addition to my theoretical work, I apply relative entropy coding algorithms to neural compression using generative models such as variational autoencoders, implicit neural representations, and diffusion models.
My personal webpage is here: https://gergely-flamich.github.io/