Data Compression with Relative Entorpy Coding


Gergely Flamich

04/02/2025

gergely-flamich.github.io

About me

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

🤔 Huge volume of data, new needs

💡ML revolution in data compression

🎯 REC: 🚀 faster and 🔒 private data compressors

🎯 Application: 📈 scalable, ⚡-efficient adaptive ML-based compressors: INRs

🎯 Practice: compress 🧬 scientific data, 🎥 video

Compression for adaptation

🤔 Can language models (ever) understand data?

💡"Occams razor:" algorithmic information theory

🎯 Theory: rule extrapolation and algorithmic causality.

🎯 Application: improved 🛠️ algorithm design, AI tools for scientists 👩‍🔬

Why Imperial?

the sponsor and host lab

Deniz Gunduz
Prof Deniz Gunduz
  • 🎯 management
  • 🤝 network
  • 🔬 collaboration

how the ICRF will benefit me

  • Develop next-gen 🗜️ not possible in 🏭
  • 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