Data Compression with Relative Entorpy Coding
Gergely Flamich
04/02/2025
gergely-flamich.github.io
Education
-
2014 - 2018
- Joint BSc in Maths & Computer Science
- Valedictorian in Computer Science
-
2018 - 2019, 2020 - now
- MPhil, graduated with commendation
- PhD, about to finish
Experience
-
- Thesis topic: data compression
- 13 papers, 10+ collabs., top ML and IT venues
- Supervised 3 MPhil theses, 2 undergraduate research projects
-
- 2022: Invented compression algorithm
- 2024: Will work on machine translation
The Proposed Project
Adaptive Compression
đ¤ Huge volume of data, new needs
đĄML revolution in data compression
đŻ REC: đ faster and đ private data compressors
đŻ Application: đ scalable, âĄ-efficient adaptive ML-based compressors: INRs
đŻ Practice: compress 𧏠scientific data, đĽ video
Compression for adaptation
đ¤ Can language models (ever) understand data?
đĄ"Occams razor:" algorithmic information theory
đŻ Theory: rule extrapolation and algorithmic causality.
đŻ Application: improved đ ď¸ algorithm design, AI tools for scientists đŠâđŹ
Why Imperial?
the sponsor and host lab
Prof Deniz Gunduz
- đŻ management
- đ¤ network
- đŹ collaboration
how the ICRF will benefit me
- Develop next-gen đď¸ not possible in đ
- Enterprise Lab
- Establish independence, smooth transition to lectureship
Envisioned Outcomes
- Demonstrated practical, real-world advantage of my compression algorithms over state-of-the-art
- Established independence, research supported by successful grant applications (e.g. UKRI FLF)
- Well on my way to commercialise my work