About

I am a postdoc working with Martin Jaggi in the School of Computer and Communication Sciences at EPFL. Previously, I completed my PhD at Stanford University, where I was fortunate to be supervised by Mert Pilanci. Before my PhD, I did my master's in computer science at the University of British Columbia, where I was advised by Mark Schmidt. My studies were supported by the National Science Foundation and by NSERC.

My research interests are in optimization for machine learning. Right now, I am focused on trying to understand deep neural networks through the lens of convex optimization.

Activities

2026

Nov. 23-27: I will give a talk at the RICAM Optimization in AI workshop as part of their special semester on AI in scientific computing.
Oct. 1: I relocated to EPFL! Here is my official EPFL profile.
June 5: We held our minisymposium on global optimization for neural networks at SIAM OP26. You can see the list of excellent speakers here: part one and part two.
March 16: I submitted my PhD dissertation!
March 1: I am very excited to announce that I will join Martin Jaggi at EPFL as a post-doc!
Feb. 27: A Library of Mirrors was accepted to SIAM SIMODS. Congratulations to the first author, Emi Zeger!
Feb. 12: I was offered an NSERC Canada Postdoctoral Research Award!
Jan. 26: I visited the Oxford Mathematical Institute and gave a talk on convex reformulations of neural networks.
Jan. 22: Our minisymposium on global optimization for neural networks was accepted at SIAM Op26. Please join us on June 5th in Edinburgh!
Jan. 15: I visited EPFL and gave a talk on level set teleportation for the MLO laboratory.
Jan. 13: I gave a talk on convex reformulations of neural networks at the Sierra team seminar at Inria Paris.

2025

Nov. 18: I successfully defended my PhD! Many, many thanks to all the mentors, collaborators, and friends who helped me reach this point.
Oct. 27-28: I will give a talk at the 2nd SIAM Northern and Central California Sectional Conference (NCC25).
Oct. 16: I was recognized as a top reviewer for NeurIPS 2025.
Sept. 9: I am searching for postdocs starting in September 2026. Please reach out if you're interested in working with me!
July 22: We received best poster at ICCOPT 2025 for our work on directional smoothness! Many thanks to the organizers and to my co-authors.
June 14: New preprint on glocal smoothness, a local smoothness assumption for comparing gradient methods.
May 25: I was recognized as a top reivewer for ICML 2025.
May 13: I started an internship at the Voleon Group in Berkeley.
May 12: I will give a talk on the solution sets of ReLU networks at SIAM NCC25.
May 5: I will present our paper on directional smoothness at ICCOPT 2025. Please come to our poster on Tuesday, July 22!
Apr. 22: I was recognized as a top reivewer for AISTATS 2025.
Feb. 11: Our work on mode connectivity in neural networks was awarded an oral at ICLR 2025. All credit goes to the first author, Sungyoon Kim!
Feb. 10: I will give a talk on level set teleportation (appearing at AISTATS 2025) at the Montréal Machine Learning and Optimization seminar.
Jan. 22: One paper accepted to AISTATS 2025 and another accepted to ICLR 2025!

2024

Nov. 25: I visited Martin Jaggi's group at EPFL to give a talk on solution sets of ReLU networks. The slides are available on Github.
Sept. 25: One paper accepted at NeurIPS 2024. Many thanks to my fantastic collaborators!
Sept. 1: I was awarded a student researcher fellowship by the France-Stanford Center!
Aug. 16: I visited MPI Intelligent Systems and the ELLIS Institute in Tübingen to give a talk on level set teleportation. The slides are available on Github.
Aug. 5: TMLR added me to their list of expert reviewers.
July 23: I was a recognized as a top reviewer for ICML 2024.
July 1 - Dec 31: I am interning with Francis Bach at Inria in Paris.
May 28 - June 4: I visited Emtiyaz Khan at RIKEN AIP in Tokyo and gave a talk on the solution set of two-layer ReLU networks.
April 30: I will intern on the Inria Sierra Team this summer. Drop me a line if you're in Paris!
April 16: I am organizing "optimization lunches" this quarter. Join our mailing list for details.
April 4: I finally cleaned up the stochastic acceleration results from my MSc thesis. See my new arXiv paper with Mert Pilanci and Mark Schmidt for details.
March 8: Three new preprints on arXiv! We develop new theory for level-set teleportation, build up a new generalization of Lipschtiz gradients called directional smoothness, and study neural networks with scalar inputs.
Jan. 23 - Feb. 2: I visited the Flatiron Institute in New York City.
Jan. 25: I gave a talk at the MML Seminar on the solution sets of ReLU networks.

2023

Nov. 27: I was recognized as a top reviewer for NeurIPS 2023.
Oct. 26: Two workshops papers accepted at NeurIPS OPT2023!
July 4: New arXiv paper on greedy rules for block-coordinate descent with affine constraints. The analysis is based on steepest descent in the 1-norm, leading to short and elegant proofs.
June 3: I will talk at the SIAM OP23 minisymposium on adaptivity in stochastic optimization.
June 2: New pre-print out! We derive the set of global optima for two-layer ReLU networks.
April. 24: Our work on solution sets and regularization paths of ReLU networks was accepted to ICML 2023! The preprint will be out soon.
April. 1: I will intern at the Flatiron Institute with Robert Gower and Alberto Bietti this summer.

2022

Nov. 15: I had two submissions accepted to the NeurIPS OPT 2022 workshop!
Nov. 11: I was recognized as a top reviewer for NeurIPS 2022.
July 11: I gave a talk at the UBC Institute for Applied Mathematics (IAM) on convex reformulations of ReLU networks.
June 19-24: I attended the International HPC Summer School in Athens.
May 15: Our submission was accepted to ICML 2022!
April 21: I was recognized as a highlighted reviewer for ICLR 2022.
February 7: New preprint out! We develop convex optimization algorithms for training shallow ReLU networks by leveraging equivalent model classes and cone decompositions.

2021

September 1: I finally moved to Stanford, CA. It's great to be here!
June 30: I was an expert reviewer and top reviewer for ICML 2021.
April 10: Mark Schmidt and Frederik Kunstner received best paper at AISTATS 2021 for their work on the convergence of expectation-maximization. Congratulations!

2020

October 20: I was ranked in the top 10% of reviewers for NeurIPS this year.
September 15: I successfully "defended" my master's thesis on first-order optimization under interpolation conditions. The thesis is available here.
September 10: I have relocated (virtually) to Stanford to begin my PhD!
June 29 - July 10: I had a lot of fun attending the virtual MLSS this year — many thanks to the organizers for their hard work! My virtual poster is public on YouTube.
June 15: New preprint is on arXiv! I had great fun helping out with the experiments for this work on implicit regularization and preconditioners in generalized linear models.
April 29: I will attend the (virtual) MLSS 2020 from 28 June to 10 July this summer.

2019

September 25: I'm organizing an overview of the interplay between optimization and generalization for UBC MLRG this semester! It includes "sharp" local minima, implicit regularization, and interpolation.
September 4: I ranked in the top 50% of reviewers for NeurIPS 2019! This was my first time reviewing; I reviewed eight papers, including two emergency reviews.
September 4: Our work on the stochastic Armijo line-search has been accepted for a poster at NeurIPS 2019!
May 1 - August 31: I'm interning with Cédric Archambeau and Matthias Seeger at Amazon Research in Berlin.

2018

September 4: Our paper on low-rank Gaussian variational inference for Bayesian neural networks was accepted at NeurIPS 2018!
June 25 - 29: I was a teaching assistant for the approximate Bayesian inference tutorial at the 2018 Data Science Summer School at École Polytechnique. The tutorial resources are here.
January 21 - June 30: I joined Emtiyaz Khan as an intern at the RIKEN Center for Advanced Intelligence Project (AIP).

Contact

My email is and my GitHub is aaronpmishkin. My CV is here.

The darkness through which we are groping is too thick for us to make any pronouncements about it; we cannot even say that it is doomed to last. — C. Lévy-Strauss