My research interests are lossless data compression, information theory, and AI. Data sources usually have some type of structure, such as graphs and sets. This means we can compress them better by taking this structure into account during compression. Unfortunately, existing methods that do this are either sub-optimal (don’t achieve the Shannon bound) or computationally intractable. I’m interested in building computationally efficient compression algorithms that can be used with deep generative models on structured data, as well as their connections to bayesian methods.

Originally, I am from Florianópolis (Brazil) but I’ve lived in New Jersey, Orlando, Toronto (now), São Paulo, as well as other smaller cities in the south of Brazil.

Previously, I interned at Facebook AI Research (FAIR) with Karen Ullrich in the summer of 2021.

I spend more time than I should on Twitter.

Check out my CV (last updated: June/2022).

I’ve joined Google AI as a Student Researcher with Lucas Theis and Johannes Ballé.

I’m accepting invitations to speak at conferences, workshops, and academic events in general.

Selected Publications and Preprints #

For a complete list, please see my Google Scholar profile.

• Data-driven Optimization for Zero-delay Lossy Source Coding with Side-Information
  Elad Domanovitz, Daniel Severo, Ashish Khisti, Wei Yu
  International Conference on Acoustics, Speech, & Signal Processing (ICASSP), 2022

• Compressing Multisets with Large Alphabets
  Daniel Severo*, James Townsend*, Ashish Khisti, Alireza Makhzani, Karen Ullrich
  Data Compression Conference (DCC), 2022
  [arXiv] [code] [tweet 1] [tweet 2] [slides]
  
  This paper was also presented as
  Your Dataset is a Multiset and You Should Compress it Like One
  Best Paper Award at NeurIPS Workshop on DGMs, 2021
  [OpenReview]

• RCAQ: Regularized Classification-Aware Quantization
  Daniel Severo, Elad Domanovitz, Ashish Khisti
  Biennial Symposium on Communications (BSC), 2021
  [arXiv] [code]

• Improving Lossless Compression Rates via Monte Carlo Bits-Back Coding
  Yangjun Ruan*, Karen Ullrich*, Daniel Severo*, James Townsend, Ashish Khisti, Arnaud Doucet, Alireza Makhzani, Chris J. Maddison
  Long talk at International Conference on Machine Learning (ICML), 2021
  [arXiv] [code] [tweet] [video]

• Predicting Multiple ICD-10 Codes from Brazilian-Portuguese Clinical Notes
  Arthur D Reys, Danilo Silva, Daniel Severo, Saulo Pedro, Marcia M Sá, Guilherme AC Salgado
  Brazilian Conference on Intelligent Systems (BRACIS), 2020
  [arXiv] [code]
  
  * Equal contribution

Awards #

Talks and Media #

• IT@UofT Source Coding with Latent Variable Models, 2021
• Two ECE grad students receive Vector Institute Scholarships in AI., 2020
• Commit 77338d2 Pursuing a Career in Data Science (pt-BR), 2020
• Hipsters #106 Cool Data Science Cases (pt-BR), 2018

You may reach me at