
Research fellow @FlatironCCM and @mcmc_stan developer. Interested in computational and applied statistics. Opinions are my own.
Joined October 2021
- Tweets197
- Following177
- Followers554
- Likes455
✨Thank you @aistats_conf for the Best Paper Award!!
📜 arxiv.org/abs/2410.11067
💡 What does VI learn and under what conditions? The answer lies in symmetry.
🤝 Honored to share this award with my colleague Lawrence Saul from @FlatironInst.
Big congrats to Charles Margossian and Lawrence Saul for winning the #AISTATS 2025 Best Paper Award!
"Variational Inference in Location-Scale Families: Exact Recovery of the Mean and Correlation Matrix"
💡We provide theory on VI's ability to recover certain statistics, despite misspecification---that is in settings where we do NOT drive the KL-divergence to 0.
👉 VI is provably good at recovering the mean and correlation.
📊 Poster below!
1
287
🙏This award is this much more meaningful to me in that it celebrates my collaboration with the amazing Lawrence Saul (users.flatironinstitute.org/…).
1
3
215
🇹🇭 Just arrived in Phuket, Thailand for #AISTATS2025.
DM if you want to connect!
📜 I'll be presenting my recent work with Lawrence on Variational Inference (arxiv.org/abs/2410.11067), which provides guarantees for VI in the presence of certain symmetries.
Finally, I invite you to check out the work by my wonderful colleague @gil2rok who will present some outstanding work on MCMC.
📜 Sampling From Multiscale Densities With Delayed Rejection Generalized Hamiltonian Monte Carlo (arxiv.org/abs/2406.02741)
📊Poster session 3
📊 I'll also present the excellent Batch, Match, and Patch VI paper (arxiv.org/abs/2410.22292) by Chirag Modi, @dianarycai and Lawrence Saul -- unfortunately neither could make it.
💡This paper showcases the elegant optimization that can be done for score-based divergences!
📊 if you want to discuss work, come find me at Poster session 1. The paper was also selected for.... an oral. 👀
113
💡In the words of the AC: "This paper provides a very elegant analysis of variational inference in location-scale families, providing much-needed insights into when and why variational methods can so often accurately obtain accurate means estimates."
4
141
Charles Margossian retweeted
Come by our #NeurIPS2024 spotlight poster on EigenVI: score-based variational inference with orthogonal function expansions!
Friday 7:30pm, East Exhibit Hall A-C #3900
Link: arxiv.org/abs/2410.24054
New paper on EigenVI! To appear at #NeurIPS2024 as a spotlight.
EigenVI is an eigenvalue-based approach to black-box VI & constructs its variational approximations from orthogonal function expansions -- inspired by solutions to the Schrödinger equation in quantum mechanics.
🇨🇦 Next stop: Vancouver for #NeurIPS2024
🏃Ready to learn, share, discuss, connect and reconnect!!
💡My favorite part: meeting researchers and exchanging ideas, so please DM if you want to chat!
📜 Presenting our latest paper on Variational Inference
Co-authors: @dianarycai, Chirag Modi, @gowerrobert, David Blei, and Lawrence Saul from @FlatironInst!
Link: arxiv.org/abs/2410.24054
Poster: Friday 7:30pm, East Exhibit Hall A-C #3900
199
Excited to present our latest paper on Variational Inference at NeurIPS'24, with @dianarycai, Chirag Modi, @gowerrobert, David Blei and Lawrence Saul from @FlatironInst!
Come see our spotlight poster Friday 7:30pm, East Exhibit Hall A-C #3900!
Fitting a multimodal target distribution with variational inference is hard. The fit and sampling often needs costly iterative methods.
Check out our spotlight NeurIPS paper where we use polynomials, score matching, and a direct solution for the fit. 1/x openreview.net/forum?id=ad97…
Charles Margossian retweeted
New paper on EigenVI! To appear at #NeurIPS2024 as a spotlight.
EigenVI is an eigenvalue-based approach to black-box VI & constructs its variational approximations from orthogonal function expansions -- inspired by solutions to the Schrödinger equation in quantum mechanics.
Recordings of StanCon '24 are now released on @mcmc_stan's youtube channel!
youtube.com/playlist?list=PL…
These talks are a wonderful snapshot of the applied and methodological work happening in the Stan, and more broadly, Bayesian community.
A blog post on what makes an MCMC algorithm GPU-friendly and more precisely, allows you to efficiently run many chains in parallel.
statmodeling.stat.columbia.e…
Charles Margossian retweeted
In case you missed it: my recent talk on the No-U-Turn Sampler is now available! I share new insights into its reversibility and mixing time. Enjoy! youtube.com/watch?v=eAVXROsy… #MCMC #mcmc_stan #pymc3 #nimble #turinglang #numpyro #tensorflow #pymc_labs #Statistics #MachineLearning
Dear colleagues, I'm in the last year of my postdoc at @FlatironInst and I'm looking for research and faculty positions in statistics, data science, and machine learning (for the summer/fall of '25).
DM me if you know of places which would be a good fit!
For details about my research interest and teaching experience: charlesm93.github.io
CV: charlesm93.github.io/files/c…











