Andrés Muñoz Medina, Ph. D.
I am a Senior Staff Research Scientist at Google. My research consists of machine learning theory and statistics, practical applications of ML algorithms to optimization problems and most recently algorithmic design of practical private learning algorithms. I have also contributed to the design of Chrome’s Privacy Sandbox targeting APIs. I obtained my Ph. D. from NYU under the advise of Mehryar Mohri.
Latest news:
- I am the general chair for WSDM 2024 – For the first time WSDM will happen in Latin America. Don’t forget to sign up!
Recent Research Work
Learning from Aggregates
Our research group is designing novel ways of learning with only semi-supervised data. Learning from label proportions allows the learner to observe only the fraction of positive and negative examples in a group. This enhances individual privacy protection while providing a way to learn general rules about the world.
Research Outputs:
Easy Learning from Label Proportions (Neurips 2023)
A Unified Analysis of Label Inference Attacks
Fast Private Training
The celebrated DP-SGD algorithm is one of the cornerstones of private ML training. However, naïve applications of this algorithm can increase memory and runtime by 100x+ making it unusable for all but the smallest models. My team constantly strives to improve this through new research on sparse DP-SGD and open sourced libraries for ghost clipping.
Research Outputs:
A Unified Fast Gradient Clipping Frameworkfor DP-SGD (Neurips 2023)
Differential Privacy Testing
Differential privacy is becoming the de-facto technology for privacy safe information sharing. Both industry and academia are deploying large differentially private systems to protect user privacy. But are they correctly implemented? Testing for differential privacy is known to be NP-hard, yet our team focuses on developing good heuristics to detect bugs on differentially private systems.
Research Outputs:
RényiTester: A Variational Approach to Testing Differential Privacy
Research in LATAM engagements
I am passionate about expanding AI research collaborations with Latin America. I partner with organizations such as Latinx in AI, CAHSI and more recently WSDM to bring Latin American researchers to top AI conferences as well as bring top AI conferences to LATAM.
Get in Touch
If you are interested in ML and privacy research or want to figure out ways of bringing LATAM to the fore-front of AI don’t hesitate to contact me.
ammedina@google.com
ammedinaml.wordpress.com 