Hanjun Dai (@hanjundai) / X

Hanjun Dai

90 posts

Hanjun Dai

@hanjundai

@meta cooking models | ex founder; prev researcher @GoogleDeepMind, @openai

Mountain View, CA

hanjun-dai.github.io

Joined September 2011

Hanjun Dai
@hanjundai
Jul 12, 2020
Autoregressive graph generation is powerful but slow. Our recent work reduces its complexity from O(V^2) to O((E+V) log V), with sublinear mem cost and training parallelism. Paper: arxiv.org/abs/2006.15502 Code: github.com/google-researc… w/ @Azade_na @liyuajia @daibond_alpha Dale
GIF
Hanjun Dai
@hanjundai
Jun 18, 2024
Our team (w/Dale, @daibond_alpha, @mengjiao_yang + others) at Google DeepMind is looking to hire. If you are interested in foundation models+decision making, and making real-world impact through Gemini and cloud solutions, please consider applying through
job-boards.greenhouse.io
DeepMind
81K
Hanjun Dai
@hanjundai
Apr 13, 2022
Our team at Google Brain (w/ Dale Schuurmans, @daibond_alpha, @rl_agent, @mengjiao_yang and many others) is hiring a SWE to work on representation learning for reasoning, search & decision making. Apply below if you are interested! careers.google.com/jobs/results/8…
Hanjun Dai
@hanjundai
Dec 12, 2023
🎲 While we are used to sample the next token from autoregressive LMs, the same model can be decoded differently (e.g., fill in the middle) if we use fancier sampling methods ---- Introducing Discs (github.com/google-researc…), a package for discrete sampling #NeurIPS2023 🧵1/5
14K
Hanjun Dai
@hanjundai
Dec 7, 2020
Learning discrete energy-based model is hard. We present ALOE, which formulates the negative sampling process as learnable local search for contrastive learning. Paper: arxiv.org/pdf/2011.05363… Code: github.com/google-researc… w/ @rishabhs @daibond_alpha @RandomlyWalking Dale [1/2]
Hanjun Dai
@hanjundai
Oct 13, 2023
Our latest work on exploring new paradigms of learning in the LLM era. TL;DR - LLMs can learn domain customized rules from demonstration data, and reuse them to answer new questions in this domain.
Zhaocheng Zhu
@zhu_zhaocheng
Oct 12, 2023
🔥 When talking about training LLMs, do you think of updating model parameters? In fact, you can use LLMs to learn a rule library. This not only improves multi-step reasoning, but also has many advantages: interpretability, transferability, and applicable to black-box LLMs. 🧵1/6
6.1K
Hanjun Dai
@hanjundai
May 22, 2023
Lots of exciting things happening in discrete spaces recently, including human/program language, proteins, comb-opt. Look forward to your participation in our #ICML2023 workshop! Please also check our website for more information
Haoran Sun
@Ha0ranSun
May 22, 2023
I am excited to announce the call for papers for our #ICML2023 workshop "Sampling and Optimization in Discrete Space" (July 29th, Hawaii🏖️). We welcome submissions covering all aspects of #discrete machine learning. Website: sods-icml2023.github.io
sods-icml2023.github.io
Sampling and Optimization in Discrete Space (SODS)
ICML 2023 Workshop
4.3K
Hanjun Dai
@hanjundai
Dec 8, 2022
Registration is free! We will cover some recent advances in reasoning over graphs, as well as hands-on learning sessions. Look forward to seeing you during this online tutorial!
Hongyu Ren
@ren_hongyu
Dec 8, 2022
Come join us for the tutorial @ Learning on Graph: Complex Reasoning over Relational Database, we will present SMORE + Scallop! @StanfordAILab @CIS_Penn @GoogleAI @LogConference Time: 9-1030am PT, Dec. 10 Webpage: snap.stanford.edu/logtutorial/ (Attendees will get a virtual cookie👇)
Hanjun Dai
@hanjundai
Jul 12, 2020
This is a new structural searching algorithm that finds the optimal chemical synthesis plans w.r.t the given cost function. Welcome to our ICML session next week if you are interested in more details! icml.cc/virtual/2020/p…
Binghong Chen
@binghong_chen
Jul 12, 2020
Our #ICML2020 paper "Retro*: Learning Retrosynthetic Planning with Neural Guided A* Search" proposed a more efficient learning-based approach to find high-quality routes for synthesizing molecules! Check out our paper and code at arxiv.org/pdf/2006.15820… github.com/binghong-ml/re…
Hanjun Dai
@hanjundai
Nov 1, 2021
SMORE(github.com/google-researc…) is actively developing and new multi-hop models and features are coming soon. Pull-requests like docs and code are welcome! Current version is a joint work w/ @ren_hongyu @daibond_alpha @xinyun_chen_ @jure @denny_zhou and Dale.
Jure Leskovec
@jure
Nov 1, 2021
Excited to share our collaboration with @GoogleAI: SMORE is a scalable knowledge graph completion and multi-hop reasoning system that scales to hundreds of millions of entities and relations. @ren_hongyu, @hanjundai, et al. arxiv.org/abs/2110.14890 github.com/google-researc…
GitHub - google-research/smore
From github.com
Hanjun Dai
@hanjundai
May 28, 2024
nice work on flow matching for categorical data! look forward to more results with large vocabularies.
Ge Liu
@GeLiuSaber
May 28, 2024
1/4 Can we unleash the power of continuous Riemannian Flow Matching on discrete generation? We introduce “Statistical Flow Matching (SFM)”, a novel FM framework on the manifold of probability measures inspired by information geometry. arxiv.org/abs/2405.16441 Kudos to@ccr_cheng
GIF
GIF
2.2K
Hanjun Dai
@hanjundai
Apr 27, 2022
If you are interested in efficient MCMC in discrete space, please join our ICLR poster session iclr.cc/virtual/2022/p… at 6:30 pm PST tonight!
Haoran Sun
@Ha0ranSun
Apr 27, 2022
How to efficiently sample from energy based models in discrete space？Check our ICLR spotlight openreview.net/pdf?id=JSR-YDI… (poster at 6:30 pm PST) for the asymptotically optimal design — we develop a multi-steps locally balanced proposal via auxiliary path. @hanjundai, Wei, Arun
Hanjun Dai
@hanjundai
Mar 9, 2023
Prune Transformers in a few minutes and save FLOPs for language models!
Azade Nova
@Azade_na
Mar 9, 2023
Check out our gradient-free structural pruning approach for large models. No need for retraining or labeled data! In just a few minutes on one GPU, reduces up to 40% of the original FLOPs at minimal loss of accuracy. w/ @hanjundai, Dale Schuurmans paper: arxiv.org/abs/2303.04185
1.4K
Hanjun Dai
@hanjundai
May 6, 2021
Welcome to our ICLR poster session today if you are interested in how a learned local search algorithm can help find better molecules!
Binghong Chen
@binghong_chen
May 6, 2021
Our #ICLR2021 paper "Molecule Optimization by Explainable Evolution" proposed an effective EM-like explainable evolutionary process for De novo molecular design. Check out our paper and code at openreview.net/pdf?id=jHefDGs… github.com/binghong-ml/Mo…