基本信息
views: 45
![](https://originalfileserver.aminer.cn/sys/aminer/icon/show-trajectory.png)
Bio
A specific direction I'm excited about is how to get powerful models to accurately report their uncertainty. Language models are surprisingly good at imitating human behavior, but frequently "hallucinate" incorrect information, making them hard to trust. How can we robustly measure the knowledge of these models, so that we know when to trust them even if we can't directly verify their reasoning? I discuss this more in my blog post on "uncertain simulators" and my recent paper "Experts Don't Cheat". I've also explored ways to summarize the uncertainty in generative models of code, described in my paper about the R-U-SURE system.
More broadly, I'm interested in methods and tools for better understanding what neural models know, how they learn it, and how they use that knowledge to make their predictions. There's so little we know about how today's models work, and I think we could learn a lot about large model behavior by studying models in controlled settings. To this end, I recently released a JAX neural network library, Penzai, that makes it easy to inspect, modify, and visualize parts of pretrained models. It's my hope that Penzai can lower the barrier of entry for research into understanding neural networks and steering them toward safe behaviors.
In the past, I have worked on generative models of discrete data structures (such as trees, sets, and graphs), theoretical analyses of self-supervised learning, a strongly-typed language (Dex) for building unconventional machine learning models, generative models for music, and many others. See my research page for more information.
More broadly, I'm interested in methods and tools for better understanding what neural models know, how they learn it, and how they use that knowledge to make their predictions. There's so little we know about how today's models work, and I think we could learn a lot about large model behavior by studying models in controlled settings. To this end, I recently released a JAX neural network library, Penzai, that makes it easy to inspect, modify, and visualize parts of pretrained models. It's my hope that Penzai can lower the barrier of entry for research into understanding neural networks and steering them toward safe behaviors.
In the past, I have worked on generative models of discrete data structures (such as trees, sets, and graphs), theoretical analyses of self-supervised learning, a strongly-typed language (Dex) for building unconventional machine learning models, generative models for music, and many others. See my research page for more information.
Research Interests
Papers共 22 篇Author StatisticsCo-AuthorSimilar Experts
By YearBy Citation主题筛选期刊级别筛选合作者筛选合作机构筛选
时间
引用量
主题
期刊级别
合作者
合作机构
arxiv(2024)
Cited0Views0Bibtex
0
0
JOURNAL OF COMPUTATIONAL PHYSICS (2024): 112624-112624
CoRR (2023)
ICML (2023): 15262-15306
Cited3Views0EIBibtex
3
0
CoRR (2023)
ICLR 2023 (2022)
Load More
Author Statistics
Co-Author
Co-Institution
D-Core
- 合作者
- 学生
- 导师
Data Disclaimer
The page data are from open Internet sources, cooperative publishers and automatic analysis results through AI technology. We do not make any commitments and guarantees for the validity, accuracy, correctness, reliability, completeness and timeliness of the page data. If you have any questions, please contact us by email: report@aminer.cn