@StormRaiser123 我的理解是在batch size不太大的情况下，大部分时间其实花在把data从显存load到cache上？计算量的增大在data比较少的情况下，对GPU不应该是瓶颈
中文
Ruosong Wang
123 posts
Ruosong Wang
@RuosongW
CS PhD Student in CMU. Interested in reinforcement learning
Pittsburgh, PA Katılım Mayıs 2019
180 Takip Edilen1K Takipçiler
Ruosong Wang retweetledi
Is a good representation sufficient for sample-efficient offline RL? In practice, offline RL can be very sensitive to distribution shift even when using features from pretrained neural networks on the same task
arxiv.org/abs/2103.04947
w/t @RuosongW, Wu & @ShamKakade6 #icml2021
English
Update: I will be a Research Scientist at Google Brain starting this fall!
Super grateful to all of the fantastic researchers I met through the interview process. Thanks for valuable career advice & research chats – I'll take these lessons with me in the next stage of my career!
English
Ruosong Wang retweetledi
More bad news from those mean theoreticians: Unlike in planning with a simulator, even a positive action gap is insufficient for efficient online RL if all one has is linear q^* realizability. Yuanhao Wang's talk is next Tuesday at the RL Theory Seminar. tinyurl.com/rmax5y74
English
Ruosong Wang retweetledi
Our next talk:
06/08: Yuanhao Wang (Princeton, @yuanhaowang3)
"An Exponential Lower Bound for Linearly-Realizable MDPs with Constant Suboptimality Gap"
For details, please see the website:
sites.google.com/view/rltheorys…
English
如果做了 weight normalization，虽然 weight 的 norm 对 forward 的结果没有影响，但在使用自适应的优化算法时 增大 norm 效果相当于减小学习率，因此对训练过程仍然有影响……我其实一直也意识到了这一点但从来没有实际观察过影响有多大……结果今天掉坑了……
🍫🩷StormRaiser🩵🌿@StormRaiser123
草
中文
@StormRaiser123 arxiv.org/abs/1910.07454 这里证明了gd+wd 对带normalization的网络来说等价于gd+exp lr 所以用了normalization之后wd是非常重要的啊
中文
我那天从头到尾给他演示的内容，他今天说一周之内复现出来，我倒要看看他能不能搞定
（我觉得悬）
🍫🩷StormRaiser🩵🌿@StormRaiser123
说好的晚上给室友X演示如何上手一个全新的项目，他跑哪去了
中文
唉，话说室友X作为一个曾经和那么多逆天的人在一起学习过的计科30的毕业生难道就一点压力都没有吗……真是看得开……问题是如果完全不想努力的话你读哪门子的博啊……
中文
Ruosong Wang retweetledi
Our next talk:
03/09: Ruosong Wang (CMU)
"What are the statistical limits of offline RL with linear function approximation?"
For details, please see the website:
sites.google.com/view/rltheorys…
English
Ruosong Wang retweetledi
Advisor abuse of power is a serious problem in academia. I know victims who are reluctant to discuss it for fear of damaging their careers, especially if they are on a visa. I'm really happy to see this MIT org pushing to do something about it. Hopefully others follow.
English
Ruosong Wang retweetledi
Alekh Agarwal, Akshay Krishnamurthy, and I finished a major tutorial on "Theoretical Foundations of Reinforcement Learning" (hunch.net/~tforl/) for FOCS (focs2020.cs.duke.edu/program/), but potentially of much broader interest. We'll be doing Q&A Friday.
English
why no one is talking about this?!?! kept me laughing for 10min and I could hardly control myself... and still it's beyond my human imagination how this whole four seasons story could've happened...
Elsie Bevan 🏴@elsie_bevan
@shoe0nhead Perhaps this is what this means?
English
Ruosong Wang retweetledi
Planning with Submodular Objective Functions: Instead of maximizing cumulative reward, maximize the objective induced by submodular function. Standard planning & submodular maximization can be viewed as special cases
arxiv.org/abs/2010.11863
w/ @RuosongW, Zhang, @dchaplot et al
English
This looks impressive
Pierre Foret@Foret_p
Introducing SAM: An easy-to-use algorithm derived by connecting PAC Bayesian bounds and geometry of the loss landscape. Achieves SOTA on benchmark image tasks (0.3% error on CIFAR10, 3.9% on CIFAR100) and drastically improves label noise robustness. arxiv.org/abs/2010.01412
English
Ruosong Wang retweetledi
Duke Comp Sci PhD candidate Hanrui Zhang receives best student paper award at 28th European Symposium on Algorithms (ESA) Sept. 7-9, 2020. A 4th-yr PhD student, Zhang's research area is Econ. & Computation; he's advised by Prof. Vincent Conitzer. Congrats! bit.ly/dukecs-ESA2020…
English