FREDERIC SALA

I study the fundamentals of data-driven systems.

I am also a research scientist at Snorkel, where we are building a data-first approach to AI.

Previously, I was a postdoc in Stanford CS, associated with the Hazy group. I completed my Ph.D. in electrical engineering at UCLA, where I worked with the LORIS and StarAI groups.

Bio | CV | Google Scholar | Twitter | Interests

• I will be joining the Computer Sciences Department at the University of Wisconsin-Madison in January 2021! Looking for students and postdocs excited to unravel the mysteries of machine learning and beyond. Contact me for more details!
• We're hosting a workshop at NeurIPS 2020 on Differential Geometry meets Deep Learning (DiffGeo4DL). Details soon!
• I will serve as an area chair for IJCAI 2021.
• Our paper on speeding up weak supervision was accepted to ICML 2020!
• Excited to share accepted papers at AISTATS 2020 (using weak supervision to improve causal inference) and ACL 2020 (how does hyperbolic geometry help us with building better knowledge bases?).
more
Our paper on multiresolution weak supervision for sequential data was accepted to NeurIPS2019.
I gave a talk for our paper on learning dependency structures for weak supervision at ICML 2019!
I gave a talk on non-Euclidean embeddings and ML at the 2019 Physics in ML as part of ml4science, hosted by Berkeley.
We presented our work on mixed-curvature geometry for improved representations at ICLR 2019!
We hosted the Graph Representation Learning Workshop at ICLR 2019!

Obtaining large amounts of labeled data is such a bottleneck that practitioners have increasingly turned to weaker forms of supervision. We study efficient algorithms for synthesizing labels from weak supervision sources with theoretical guarantees, an efficient way to learn the structure of a model of such sources, and new ways to tackle labeling data for large-scale video and other applications.

Modern ML methods require first embedding data into a continuous space-traditionally Euclidean space. However, the structure of data makes Euclidean space unsuitable for many types of data (like hierarchies!) My work shows that non-Euclidean spaces like hyperbolic space (and other manifolds!) are more suitable for embeddings and study the limits and tradeoffs of these techniques.

What is the least amount of information we must exchange to synchronize between two versions of a file, or to reconstruct a core piece of data from noisy samples? My work studies bounds and algorithms for these techniques.

Modern memories offer speed and efficiency but suffer from physical limitations that lead to errors and corruption. New reliability and error-correcting techniques are critical to the future of these devices. My work develops new data representations, new coding techniques, and how to make algorithms more robust. I also study theoretical frameworks to evaluate broad ranges of ECCs.

• Top Reviewer NeurIPS, 2018, 2019
• Outstanding Ph.D. Dissertation Award, UCLA Department of Electrical Engineering (Signals & Systems Track) (2017)
• UCLA Dissertation Year Fellowship (2015-2016)
• Qualcomm Innovation Fellowship Finalist (2015)
• Edward K. Rice Outstanding Masters Student Award UCLA Henry Samueli School of Engineering & Applied Science (2013)
• Outstanding M.S. Thesis Award, UCLA Department of Electrical Engineering (Signals & Systems Track) (2013)
• National Science Foundation Graduate Research Fellowship (NSF GRFP) (2012-2015)