About me
Research interests
My central focus of research is in theoretical understanding of statistical machine learning, particularly from the following perspectives.
- Learning theory of loss functions: Loss functions are interesting because they characterize a large portion of task properties such as adversarial robustness (COLT2020) and imbalancedness (AISTATS2020, AISTATS2021) Some task properties can be attained simultaneously by proper losses (COLT2023).
- Evaluation metrics of predictions and representations. Recently, I am interested in how it is possible to learn good representations via similarity in light of a downstream task (ICML2018, AISTATS2022, ICML2022).
You may have a look at the slides of my past (and slightly outdated…) talks such as this to see my tastes.
News
For prospective students: If you are willing to join our team or visit me to initiate any research collaboration, please contact me with your research interests (ideally with your previous research achievements) and relevance to my research. There would be availablity depending on time.
- Jul 10, 2024: We are going to organize ECCV 2024 Workshop “Dark Side of GenAI and Beyond” at Milano, Italy, on Septempber 29. We highly encourage those who have ideas on possible social effects of generative/foundational models and their remedies (regardless of the maturity) to give a poster talk. The deadline is July 31.
- May 9, 2024: Our paper “Online Structured Prediction with Fenchel–Young Losses and Improved Surrogate Regret for Online Multiclass Classification with Logistic Loss” has been accepted by COLT2024.
- May 2, 2024: Our paper “Self-attention Networks Localize When QK-eigenspectrum Concentrates” has been accepted by ICML2024.
- Jan 20, 2024: Our paper “Fast 1-Wasserstein Distance Approximations Using Greedy Strategies” has been accepted by AISTATS2024.
- Dec 6, 2023: I’m going to attend NeurIPS2023 in New Orleans next week to present our work “Beyond Exponential Graph: Communication-Efficient Topologies for Decentralized Learning via Finite-time Convergence”: propose Base-k graph, which is a finite graph sequence for any number of nodes and facilitates better communication convergence in decentralized learning.
- Dec 5, 2023: My presentation at IBIS2023 was selected as the presentation award finalist!
- May 15, 2023: Our paper “Proper Losses, Moduli of Convexity, and Surrogate Regret Bounds” has been accepted by COLT2023. We establish a connection between the surrogate regret bound of a proper loss and the moduli of convexity of its (generalized) entropy.
Show more
- May 3, 2023: Our paper “Unbalanced Optimal Transport for Unbalanced Word Alignment” has been accepted by ACL2023. We show the effectiveness of unbalanced optimal transport in monolingual word alignment tasks, where the null alignment ratio is high.
- Mar 16, 2023: I received Funai Information Technology Award for Young Researchers (船井研究奨励賞) for my PhD work.
- Dec 21, 2022: Our presentation at IBIS2022 got the presentation award!
- Nov 11, 2022: Our paper “Sparse Regularized Optimal Transport with Deformed q-Entropy” has been accepted by Entropy. We show a possible formulation of sparse optimal transport via q-exponential distributions in Tsallis statistics.
- Sep 19, 2022: Our paper “Approximating 1-Wasserstein Distance with Trees” has been accepted by Transactions on Machine Learning Research (TMLR). For any ground metric of 1-Wasserstein distance, we proposed an approximation method with a tree metric.
- Sep 17, 2022: Our paper “Robust Computation of Optimal Transport by β-potential Regularization” has been accepted by ACML2022.
- Aug 23, 2022: Our monograph “Machine Learning from Weak Supervision: An Empirical Risk Minimization Approach” has been published from MIT press.
- May 16, 2022: Our paper “On the Surrogate Gap between Contrastive and Supervised Losses” has been accepted by ICML2022. We improve upper and lower bounds for the gap between contrastive and supervised losses and claim that larger negative samples are good for downstream classification. The earlier version is available here.
- May 9, 2022: Our AISTATS2022 paper “Pairwise Supervision Can Provably Elicit a Decision Boundary” has appeared in the proceedings (link).
- Apr 1, 2022: I have joined Kyoto University as an assistant professor. Feel free to visit Kyoto and contact me.
- Mar 24, 2022: I finished my three year PhD in computer science and nine years life in the University of Tokyo. Also, I was fortunate to have an opportunity to be a representative graduate at the diploma presentation ceremony to make an address in the ceremony. My student life has been supported by so many great friends not only in Tokyo but also in other cities in Japan and even in overseas. I would like to appreciate everyone who has been with me!
- Jan 19, 2022: Our paper “Pairwise Supervision Can Provably Elicit a Decision Boundary” has been accepted by AISTATS2022. We elucidated that pairwise supervision (i.e., information indicating whether two input vectors belong to the same underlying class) is sufficient to recover a binary decision boundary. The latest version is available here (updated on Mar 3).
- Jun 21, 2021: Our paper “Learning from Noisy Similar and Dissimilar Data” has been accepted by ECMLPKDD2021.
- May 17, 2021: We have publicized a corrigendum to our COLT2020 paper. The definition of calibrated losses is corrected and the proofs of our main results are modified.
- Jan 23, 2021: Our paper “Fenchel-Young Losses with Skewed Entropies for Class-posterior Probability Estimation” has been accepted by AISTATS2021!
- Jan 8, 2021: Our presentation at IBIS2020 got the best presentation award (1st place out of 116 presentations)!
Upcoming travels
- Jul 22-27: Vienna (@ICML)
- Jul 28-31: Tuebingen (@Univ. of Tuebingen)
- Aug 9-10: Tokyo (@Univ. of Tokyo)
- Aug 19-21: Yamagata (@JCCA-DMIA-2024)
- Sep 28-Oct 1 (TBD): Milano (@ECCV)
- Nov 4-7: Omiya (@IBIS)