Zheng Chen (@chenzch) header
Zheng Chen (@chenzch) profile picture

Research Asst. Prof @osaka_univ, ビール好き.

Joined February 2017
Photo by Zheng Chen (@chenzch)Photo by Zheng Chen (@chenzch)Photo by Zheng Chen (@chenzch)Photo by Zheng Chen (@chenzch)Photo by Zheng Chen (@chenzch)Photo by Zheng Chen (@chenzch)
Zheng Chen retweeted
🧠⚡ Excited to release EpiGraph! Led by PhD student Yuyang Dai, in collaboration with The University of Osaka (Dr. Zheng Chen, Dr. Yasuko Matsubara, Dr. Yasushi Sakurai) and University of Illinois Urbana-Champaign (Jathurshan Pradeepkumar, Dr. Jimeng Sun). 🎉
5
22
77
7,559
Zheng Chen retweeted
The takeaway: structured epilepsy knowledge substantially boosts evidence-grounded clinical reasoning! 🔌 Plug-and-play — run EpiBench with your own model, retriever & prompts 🌐 Interactive project page — explore the KG, click nodes, inspect evidence paths 🆓 Fully open-source
1
2
145
Zheng Chen retweeted
Are you curious about how SE researchers apply Reinforcement Learning (RL) in their SE tasks? We recently conducted a survey of RL4SE, including the trends, research topics, used RL algorithms, model design and optimization, and more! Please check it! arxiv.org/abs/2507.12483
2
2
4
468
Rule-based rewards (RBRs) use model to provide RL signals based on a set of safety rubrics, making it easier to adapt to changing safety policies wo/ heavy dependency on human data. It also enables us to look at safety and capability in a more unified lens as a more capable grader model gives us higher quality RL signals.
12
42
3
301
145,847
Zheng Chen retweeted
Checkout a new ICML paper from our group led by @ori_press! 🚨 We show why entropy minimization improves the accuracy of classifiers on input data, and how it could be used to estimate the accuracy of a classifier on an arbitrary dataset, without labels. Detailed🧵👇
Entropy minimization is often used to increase the accuracy of models on unlabeled data, but it isn’t clear why it works. In our new ICML paper, we show that it clusters the embeddings of its inputs. With @ziv_ravid, @ylecun, @MatthiasBethge arxiv.org/pdf/2405.05012 1/5 🧵👇
3
16
2,278
Zheng Chen retweeted
Had a wonderful time meeting and discussing with Emanuele Rossi (@emaros96) at #ICML2023 @icml_compbio, the author of Feature Propagation, who truly inspired our work in single-cell RNA-seq.
1
16
1,020
Zheng Chen retweeted
If you missed Durk’s talk at SPIGM @ICML2023, check the recording, slides, and the latest version of our paper under this thread! 😃👇
Thanks to the SPIGM organizers and those who attended! If you're an ICML registree but missed my talk, you can find it at 2:21:41 under this link: icml.cc/virtual/2023/worksho… Our updated paper (including new results) will appear on arXiv on Aug 1. I'll also post the slides then.
7
47
12,122
Zheng Chen retweeted
Large language models (LMs) rely heavily on training data quality. How do we best select training data for good downstream model performance across tasks? Introducing 🍳Skill-It: a data-driven framework for understanding and training LMs! Paper: arxiv.org/abs/2307.14430 1/13
6
122
13
447
111,301
🔥Hot off the press🔥 Preprint of our new paper “Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting” on time series forecasting using ✨unconditional✨ diffusion models is available now. 🧵 arxiv.org/abs/2307.11494
1
7
1
39
7,672
Zheng Chen retweeted
This is a pretty cool paper. It first introduces a new multimodal biomedical benchmark with 14 different tasks. It then presents a proof of concept for a generalist biomedical AI system called Med-PaLM Multimodal. It supports different types of biomedical data like clinical text, imaging, and genomics. "In a side-by-side ranking on 246 retrospective chest X-rays, clinicians express a pairwise preference for Med-PaLM M reports over those produced by radiologists in up to 40.50% of cases, suggesting potential clinical utility." Multimodal systems keep making progress. I believe it's possible that we might end up with all these general-purpose AI systems that cater to different domains. It would be cool to incorporate these as components to more general AI systems based on needs and problems. As an example, curious how this could be combined with Galactica for instance for more research-related tasks. There is a ton of research that could be done on the interplay of LLMs - not only for evaluation but also for increasing capability. There is also a lot of work to be done to better evaluate these systems too. The proposed benchmark is a good start but for a domain like health, we need a pretty robust set of tools and benchmarks for assessing the robustness, safety, and quality of models. (paper in the replies)
3
114
6
574
147,542
The 𝗪𝗮𝘀𝘀𝗲𝗿𝘀𝘁𝗲𝗶𝗻 𝗺𝗲𝘁𝗿𝗶𝗰 is a natural way to compare the probability distributions of two random variables, first defined by Leonid Kantorovich (1939) in the context of optimal transport planning of goods. It is now used more and more in machine learning.
2
68
308
25,756
Zheng Chen retweeted
Congratulations Neuralink team!
We are excited to share that we have received the FDA’s approval to launch our first-in-human clinical study! This is the result of incredible work by the Neuralink team in close collaboration with the FDA and represents an important first step that will one day allow our technology to help many people. Recruitment is not yet open for our clinical trial. We’ll announce more information on this soon!
9,091
10,177
1,624
103,020
29,296,486
If new with graph learning, here is a warm-up notebook to learn to use graphs: • Build graph w/ features and compute basic message-passing function w/ @DGLGraph • Convert into graph formats w/ DGL, NetworkX, dense/sparse PyTorch • Visualize graph Code: github.com/xbresson/CS6208_2…
4
93
2
355
37,468
Readers asked if I can share the slides of my 2008 lecture at Georgia Tech: youtube.com/watch?v=RPgvfSeQ… Yes, I've found them: web.cs.ucla.edu/~kaoru/ppt.h… I love the lecture so much, that I'm going to use same slide next week at Rand.
2
5
1
48
12,058
Zheng Chen retweeted
Holy Mackerel! A reader sent me a video of a talk I gave 15 years ago and, by gosh, it is so much better than the talks I give today, that I can't resist sharing it: youtube.com/watch?v=RPgvfSeQ… It doesn't cover transportability, true, but the conceptual framework is crystal clear.
3
10
83
11,809
Diffusion Models are powerful for decision making: 💡 Provide an alternative to RL by stitching trajectories without dynamic programming! 💡Compose skills without hand-design of heirarchies or task-planners!! Presenting Decision Diffusers at #ICLR2023: anuragajay.github.io/decisio…
3
41
1
272
32,504
Embedding quality is the key to distance-based OOD detection. Here is a more elegant alternative to off-the-shelf contrastive loss. Sharing our #ICLR2023 paper “How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection?”. 1/n Arxiv: arxiv.org/abs/2203.04450
3
45
1
217
30,199
Zheng Chen retweeted
Very cool and thoughtful. There is clearly great potential to using LLM in CI tasks, possibly leading to a new hierarchy and a new ladder, where the input is not "data", nor "experiments", but "texts about data and experiments".
New paper: On the unreasonable effectiveness of LLMs for causal inference. GPT4 achieves new SoTA on a wide range of causal tasks: graph discovery (97%, 13 pts gain), counterfactual reasoning (92%, 20 pts gain) & actual causality. How is this possible?🧵 arxiv.org/abs/2305.00050
9
38
4
196
50,585