publications

2026

1. AISTATS

   Robust Federated Clustering under Heterogeneity and Adversaries

   Martín Bravo, and Sebastian Dalleiger

   In Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, 2026

   Abs Bib HTML

   Clustering distributed and private data is an increasingly important task across domains that handle sensitive information, such as life sciences and clinical research. In federated settings, clustering faces three challenges: heterogeneous client data distributions, adversarial behavior, and strict privacy requirements. Existing approaches often exhibit significant performance degradation under these conditions and fail to return accurate solutions. To overcome these limitations, we introduce a novel federated clustering algorithm that combines client-level differential privacy with Byzantine-robust aggregation at the server, based on a novel efficient and robust clustering procedure. Our method comes with theoretical robustness guarantees, and through extensive experiments on synthetic and real-world data, we demonstrate that it produces high-quality clusters in just a few communication rounds, even in scenarios where state-of-the-art methods fail.

   @inproceedings{bravo2026robustfederated,
     title = {Robust Federated Clustering under Heterogeneity and Adversaries},
     author = {Bravo, Martín and Dalleiger, Sebastian},
     booktitle = {Proceedings of The 29th International Conference on Artificial Intelligence and Statistics},
     year = {2026},
   }

2. Thesis

   Lossless Model Compression

   Martín Eduardo Bravo Díaz

   University of Chile, 2026

   Advisors: Andrés Abeliuk and Gonzalo Navarro

   Abs Bib HTML PDF

   Foundation models have grown substantially in both capability and computational requirements, making the storage and transfer of their weights increasingly costly. This thesis presents NeuralZip, a lossless compressor for foundation-model weights inspired by compact data structures. NeuralZip exploits blockwise heterogeneity and local exponent redundancy by clustering tensors with similar exponent distributions, reusing Huffman codes, and selecting packed-exponent representations when beneficial, while preserving bitwise-exact reconstruction. Across language, vision, diffusion, and mixture-of-experts models, NeuralZip encodes faster than ZipNN in 26 of 34 evaluated model artifacts, with speedups of up to 21.51 times.

   @mastersthesis{bravo2026lossless,
     title = {Lossless Model Compression},
     author = {Bravo Díaz, Martín Eduardo},
     school = {University of Chile},
     year = {2026},
     note = {Advisors: Andrés Abeliuk and Gonzalo Navarro},
   }