Tomorrow, I will present this work about the Strong Lottery Ticket Hypothesis at the compression workshop (West Meeting Room 211-214) of #NeurIPS2024 😁 Let's talk with me about SLTH💪 arxiv.org/abs/2402.14029

Finally, our paper about Strong Lottery Tickets within Frozen Networks has been accepted & published at #TMLR! I would like to thank the equally contributed authors, Chijiwa-san (Daiki Chijiwa) and López-san, and all the co-authors for their cooperation! openreview.net/forum?id=xpnPY…

Young researchers selected for DC1 are now publicly announced! I will work on the research topics titled "Construction of a resource-saving DNN training method based on optimal selection of neuron connections and its theoretical analysis." jsps.go.jp/j-pd/pd_saiyoi…

We are holding a machine learning workshop next week from June 25-27. Talks are streamed and you can find information and how to register at: tinyurl.com/y7wzrfb4 Many great speakers! I will also talk about my recent work which I'm pretty excited about! The Bayes Duality Project

大規模グラフニューラルネットワーク推論性能の飛躍的向上 isct.ac.jp/ja/news/vun1gp… #ScienceTokyo

明後日京都で開催されるDCフレンドシップミーティングに参加します💪

Excited to share that our work with NTT on the Strong Lottery Ticket Hypothesis for attention mechanisms has been accepted to the 3rd Workshop on High-Dimensional Learning Dynamics (HiLD) at #ICML2025! Looking forward to the discussions at the venue!👋😁 openreview.net/forum?id=gB1LZ…

#学振 🙌😄

文脈内学習の状況にて，Transformerはsoftmax注意によって「テスト時に」特徴学習ができることを示しました．さらに，そのテスト時の学習複雑さは情報理論的下限に近いレートを達成し，「生成指数」と呼ばれる量で特徴づけられることを示しました．ICML2025で発表します． x.gd/WXBCy

super interesting👀

Thrilled to announce that my JASSO scholarship has been fully waived in recognition of my academic excellence! 🎉 I’m truly thankful for this opportunity and support 🙏

Totally forgot to mention this, but I was selected as an outstanding master's student when I graduated last year! 🎉 Apparently, I was at the top of my department too! I'm so happy and grateful. I’ll keep working hard on my research from here on out 💪

他分野への数学の先駆けた応用には，異質分野の勉強と研究対象の数理化＋数学解析が必要で，準備に時間がかかると思います。甘利先生のご指摘の通り，流行や短期的でない良い研究を自由に続けられる環境整備は重要です。 甘利俊一が懸念する日本の研究環境 文春オンライン bunshun.jp/articles/-/810…

第28回情報論的学習理論ワークショップ（IBIS2025）では、11月11日 (火) に社会人学生を含む学生を対象に、若手交流企画を実施予定（詳細は近日公開）です。学生の皆さんの参加をお待ちしています。 IBIS2025本プログラム (11月12～15日) の前日であることにご注意ください。 ibisml.org/ibis2025/newin…

This is not a particularly good take and is indicative of a fundamental misunderstanding of what a top-tier technical college education is suppose to offer. Preparing to understand modern AI as a Harvard or Stanford undergrad is not about learning "prompt engineering", vibe

Absolutely the right take. People have wrongly complained about places like Harvard CS having “too much theory” forever. Turns out if you get the theory it’s not that hard to apply it. The reverse is not true. That + a culture of students working on side projects with their

We’ve found a ton of value hiring folks with strong theory backgrounds with little to no production ML experience. One of our members of technical staff got his phd in pure math/the geometry of black holes and had no prior ML experience. Within days of hiring him we released our

先日公開したプレプリントにて、言語モデルの推論中に『トークン語彙集合を自由自在に縮小』できる新たなアルゴリズムを導出しました！ この手法により、トークナイザの異なるLLM同士でも互いの次トークン分布を『共通語彙集合』上に縮小させることで、精度劣化なく連携可能になりました。

Yay! 😆 Our paper has been accepted to AAAI '26! 🥳 We prove that SLTs exist for MHAs (& transformers): a randomly-weighted MHA contains an SLT that approximates an arbitrary MHA, and this SLT existence does not depend on sequence length! #AAAI26 arxiv.org/abs/2511.04217

私と NTT 千々和さん (Daiki Chijiwa) との共同研究が、AAAI'26 に採択されました！ 本研究では、強い宝くじ (高精度乱数サブネット) が MHA 及び Transformer アーキテクチャに含まれていること、更にその存在性が近似対象の QKV 次元および入力シーケンス長に依存しないことを理論的に示しています。