About me
🔎 For students
I'm open for accepting PhD students and interns at ISM. Feel free to reach out to me. Here are some of my favorite recent papers, but I'm more broadly interested in learning theory. More info can be found here.
Research interests
- Learning theory
- classification-calibrated losses, proper scoring rule, property elicitation
- class probability estimation, probability calibration
- learning dynamics, gradient descent
- convex analysis, information geometry
- Representation learning
- contrastive learning
- robust learning
- Online convex optimization
News
- Sep 25, 2026: Our four papers are accepted by NeurIPS2026:
- “Flow Matching from Viewpoint of Proximal Operators”, led by Kenji; by rewriting OT-CFM as a proximal step of the Brenier potential, we can analyze the dynamics stability at the terminal phase.
- “Denoise First, Orthogonalize Later: Understanding Momentum in Muon via Spectral Filtering”, led by Xianliang and Zihan; momentum in Muon plays a crucial role to identify the gradient signal space, enhancing the benefit of the subsequent polar factor.
- (spotlight) “Toward Minimal-dimensional Convex Calibrated Surrogate Losses for Classification with Rejection”, led by Yuzhou; improves the convex calibration dimension (CCdim) of the 0-1-c loss (used for classification with rejection) to 2, counterarguing the commonly-believed CCdim > log(K) with K = # of classes.
- (spotlight) “Non-asymptotic Implicit Bias of Logistic Regression at Early-stage Gradient Descent Dynamics”, my solo paper; the first non-asymptotic proof of (weakly) max-margin bias, explaining empirically observed fast dynamics.
- Sep 3, 2026: I’m delighted to host Mohit as a postdoc researcher to our team!
- May 12, 2026: Three papers are accepted by ICML2026: Muon’s dynamical analysis on anisotropic covariance (led by Guillaume from RIKEN AIP), dynamic regret analysis of online structured prediction (led by Shinsaku from CyberAgent), and distortion-based analysis of RLHF under distribution shift (led by Kazusato from UCBerkeley).
- Jan 23, 2026: Our paper “Brenier isotonic regression” is accepted by AISTATS2026. We propose new isotonic regression with multi-dimensional outputs satisfying cyclic monotonicity based on the optimal transport theory.
- Nov 14, 2025: Our work on any-stepsize gradient descent convergence got IBIS2025 excellent presentation award!
- Oct 1, 2025: Japanese translation of Kevin Murphy’s textbook “Probabilistic Machine Learning: An Introduction” will be published soon from Asakura Publishing [link (vol 1)][link (vol 2)].
- (archived)
Upcoming travels
Drop me a line; I’m open to talking in-person.
- Oct 26-29: Omiya, Japan (IBIS)
- Nov 25-27: Hakata, Japan (PRESTO meeting)
- Dec 6-12: Sydney, Australia (NeurIPS, TBD)
- Dec 15-18: Split, Croatia (ICSDS)
- Feb 15-18, 2027: Delhi, India (TBD)
- Mar 1-5, 2027: Hangzhou, China (TBD)
- Mar 8-12, 2027: Okinawa, Japan (MLSS, TBD)
- (archived)