Zexuan Hao

PhD, Tsinghua University. Neuroimaging, machine learning, statistical tests, brain injury.

[AYi @AYi_AInotes]

全网都在吹的LeCun新论文，90%的解读都是错的。

他们说生成式AI是死路，说过去三年花的几百亿全白费了，说15M参数的小模型就能吊打万亿大模型。

这些全是营销号的夸张，
我觉得这篇论文的真正分量比他们吹的还要重。

Yann LeCun团队这次解决了JEPA困扰了好几年的表征坍缩问题。

以前的世界模型，学着学着就会把狗车人都压成一模一样的向量，什么都学不到。

这次他们只加了一个极其优雅的数学正则化器SIGReg，
没有复杂的trick 和六个超参数要调，训练稳得离谱。

单张GPU几个小时就能训完，在机器人控制任务上，规划速度比巨型世界模型快48倍，成功率还更高。

最厉害的是它的隐空间里天然就编码了物理规律。
不用教，它自己就知道物体不能瞬移，知道速度和位置的关系。
能瞬间检测出物理上不可能发生的事。

这不是啥范式革命，也不会让GPT和Claude明天就死掉。
语言和创意生成，依然是自回归大模型的天下。

但它打开了一扇全新的门，
原来懂物理不需要万亿参数，不需要云端超算。
原来世界模型可以小到跑在机器人的本地芯片上。

过去三年，整个行业都在一条路上狂奔，堆参数，堆算力，堆数据。
所有人都以为只要足够大，就能懂世界。

现在我们终于知道，还有另一条路，一条更高效，更优雅，更接近真实世界运行方式的路。

生成式AI不会死，
但未来的智能体不会是只会聊天的大模型，
它会是一个懂物理的小世界模型，加一个大语言接口，

这才是这篇论文真正的意义所在吧 hhh

[Daniël Lakens @lakens]

Finally got around to incorporating all Github feedback on my open access textbook, making around 20 minor improvements, and updating some references. 17 chapters of state of the art stats and methods education, freely available for any course you teach. https://t.co/FdQIS63fVd

[Christoph Molnar 🦋 christophmolnar.bsky.social @ChristophMolnar]

Tons of ways to "quantify uncertainty", but we don't talk enough about sources of uncertainty for ML models.

Is it uncertainty from model bias? Missing data or features?

Bookmark this: Sources of Uncertainty in Machine Learning -- A Statisticians' View

https://t.co/ciGY0OL0FG

[Alberto Montanari @almontanari]

A new chapter for Bluecat: uncertainty estimation for multimodel predictions, with open source software in R and Python, along with rigorous testing of the estimated uncertainty. Extremely simple and reliable, try it!!
https://t.co/6su5sbDqtS

[Statistics (Machine Learning) Papers @StatsPapers]

Do you understand epistemic uncertainty? Think again! Rigorous frequentist epistemic uncertainty estimation in regression. https://t.co/5gHWF61Lci

[Sean J. Taylor @seanjtaylor]

I came across this really awesome explanation and comparison of a variety of methods to estimate predictive intervals from neural networks (in PyTorch). A great starting point if you’re thinking about how to add uncertainty to your model. https://t.co/cpdUfP3dXh

[Paul Thompson @PTenigma]

Cool paper out tonight on uncertainty estimation in generative AI [1]. You can use generative AI in medical imaging to flag anomalies, and enhance images - but wouldn't you like it to tell you how confident it is in its prediction?
💡The paper enhances a VAE (variational auto encoder) with a SLU layer (Stochastic Laplace Uncertainty) - after the VAE makes a prediction (like generating an image or flagging an anomaly), SLU uses a Laplace approximation to figure out how confident the model is about that prediction
💡The Laplace approximation looks at the curvature of the model’s loss function (how the error changes as you alter the parameters). If the loss function is flat (low curvature), the model is less confident, and if it's steep (high curvature), the model is more confident. By calculating this, SLU figures out how confident the model is in its prediction.
Bonus Q: can you see a connection to 2nd-order Tweedie's formula? **answer below in comments, +link to the paper

[Decision Neuroscience and Neuroeconomics Papers @decisionneurop]

Uncertainty estimation with prediction-error circuits https://t.co/myh2opCo2p

[Bratislav Misic @misicbata]

Integrating brainstem and cortical functional architectures | https://t.co/hmGC40nJq2

led by @JustineYHansen in @NatureNeuro

400 cortical regions + 58 brainstem nuclei. Lesgo 🤘🕶️⤵️

[Thomas Yeo @bttyeo]

Backlog in announcing preprints, but here's goes! Preprint #3 by @anlijuncn https://t.co/b6Qh6L3fuv proposes DeepResBat, a deep learning alternative to ComBat that harmonizes MRI while accounting for covariates. We show large improvements over classical & deep learning baselines

[Juan Eugenio Iglesias @JuanEugenioIgl1]

We have updated NextBrain with:
- new atlas with improved brainstem.
- updated segmentation code.
- new labeled ex vivo dataset.
- new preprint.
Please see the project website for links and resources:
https://t.co/736q8mrfJo

[Paul Thompson @PTenigma]

WOW - @OHBM just put about 800 videos online including tutorials and symposia [1,2] from many years of past brain mapping meetings! https://t.co/8sY1np3bgd
[1] Some were there before but it looks like entire conferences were just added
[2] I am not involved with this, just appreciative - thanks to @StephForkel @sofievalk Kevin Sitek, Alfie Wearn

[Chuck Lynch @cl9681]

Happy to share that our work applying precision mapping to individuals with depression sampled longitudinally over an extended period has been published this week in @nature. Brief recap of the main findings and their potential implications below.

https://t.co/ZbTj6x7eyV

[Chris Chatham @chchatham]

I'm literate in family-wise error rates, multiplicity correction, closed testing procedures & 'alpha spending'; but I've never come across the complementary notion, of family-wise power. Am I missing a term of art , or am I confused & this notion is somehow incoherent ('MNAR')?

[Jacobo Sitt @jdsitt]

I am thrilled to see this paper out after many years of work. It was an honor to be part of this group of leaders in the field, and it was a seminal experience that fast-forwarded my scientific career.
Thank you, Niko Schiff, for leading this initiative.
https://t.co/9nXaixd69y

[Matt Rosenblatt @RosenblattMatt]

Curious about power in the external validation of brain-phenotype predictive models? Our paper "Power and reproducibility in the external validation of brain-phenotype predictions" was just published in @NatureHumBehav https://t.co/MfcqvdL5YL

[Zhiqiang Sha (@zhiqiangsha.bsky.social) @sha_zhiqiang]

Brain-wide functional connectivity artifactually inflates throughout functional magnetic resonance imaging scans | Nature Human Behaviour
https://t.co/Nh19bD7fY7

[Allen Downey @AllenDowney]

On Reddit's statistics forum, the most common question is "What test should I use?"
My answer, from 2011, is "There is only one test"

https://t.co/J5Ar4olekz

[Zexuan Hao @yihao1815]

@sNeuroble At least they were honest😂

[Imaging Neuroscience @ImagingNeurosci]

New paper in Imaging Neuroscience by Chetan Gohil, Mark W. Woolrich, et al:

Dynamic network analysis of electrophysiological task data

https://t.co/sMV3q5ET5u