Jam

Deep learner

[Roan @RohOnChain]

USC mathematicians just published the most dangerous quant paper of the year.

THE MATH BEHIND HOW INSIDERS BEAT THE MARKET WITHOUT GETTING CAUGHT.

This paper will teach you to detect smart money moving before announcements. 43 pages of pure game theory. Bookmark now.

[Lei Li @_TobiasLee]

镜像的问题是如何写论文，AI 可以解决 writing，但不能解决逻辑。
每次都会给 junior 的同学推荐清华刘洋老师的 slides:
https://t.co/TJgrEsLoJO
我自己也是常看常新，收获非常大

[deep Manifold @BetaTomorrow]

https://t.co/K9ib2Bmg0b

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

This video by @jbhuang0604 is a compact but very informative dive into the progress of self-supervised learning over the past few decades.

from IMAX in 1992

covering methods like MoCo, SimCLR, DINO, BYOL, MAE

all the way up to LeJEPA in 2025

Highly recommend watching!

[恒星 @vintcessun]

时间序列异常检测一直有个恼人的空档：算法丢来一个分数，却从不告诉你“为什么这里异常”？没有解释，用户只能干瞪眼，信任和诊断都无从谈起。ProtoX-AD 终于打破了这层窗户纸——它在自监督分类框架里嵌入可解释的原型向量层，每个原型对应一种变换模式或异常特征。训练时让正常样本聚到对应原型附近，异常样本远离所有原型；推理时分类误差用来检测，同时通过原型相似度直接告诉你异常是“局部突变”还是“趋势偏移”。检测精度不牺牲，解释还带着语义。边界是原型需预设，但首次补齐了可解释性这块关键缺失。
https://t.co/5D59ohfS6J

[Lucas Beyer (bl16) @giffmana]

LLM community slowly rediscovering what we in vision found out over half a decade ago. MY SCHMIDHUBER MOMENT IS COMING!

Source: S4L paper where i tuned the most sota 10% and 1% ImageNet baselines ever, by far.
https://t.co/Cj10TYvpOP

[Xavier Bresson @xbresson]

What are the training datasets driving AI for materials science?

As I am updating my GraphML course, I prepared a series of slides introducing the datasets in the field, with short summaries and links to the papers and datasets.

Sharing them below:
https://t.co/TJMRYpexiT

[paperpaper @paperpaper886]

最近在带入组的本科实习生，发现怎么读论文其实是科研训练里最容易被忽略的一步。
推荐一篇每个科研新人都该读的经典短文：S. Keshav 的 How to Read a Paper。

文章提出了非常实用的“三遍读论文法”：
第一遍，5 到 10 分钟快速扫读：标题、摘要、引言、章节标题、结论和参考文献。
目标是回答 5C：
Category, Context, Correctness, Contributions, Clarity。
也就是判断这篇论文是什么、和谁相关、假设是否合理、贡献是什么、写得清不清楚。
第二遍，认真读论文主线，但先跳过证明细节。重点看图表、实验设置、结果是否清楚、引用了哪些关键工作。
第三遍才进入深度理解：尝试像复现一样重建作者的思路，检查假设、方法、创新点和潜在漏洞。

放在今天看，这个方法和 AI 辅助读论文其实很契合。
第一遍可以让 AI 帮忙快速总结论文的研究问题、核心贡献和主要结论，但自己一定要判断这篇文章是否真的值得继续读。
第二遍可以让 AI 帮忙解释方法、实验设置、图表和不熟悉的概念，但不能只看 AI 总结。关键图表、实验设计和结果数字一定要回到原文核对。
第三遍可以让 AI 扮演 reviewer，帮你追问：这篇文章的假设是否成立？实验是否支持结论？有没有 missing baseline？有没有潜在的数据泄漏、评价偏差或过度 claim？

读论文不是“读完”就行。真正重要的是知道什么时候快速跳过，什么时候认真理解。
尤其在 AI 工具越来越强的情况下，科研新人更需要训练自己的判断力。
AI 可以帮你压缩信息，但不能替你决定一篇论文是否重要、是否可信、是否值得借鉴。

https://t.co/8gUc4HbLwR

[恒星 @vintcessun]

99%的输入被噪声污染，神经网络仍能分类，这反直觉吗？它没有“记住”样本，而是偷偷学会了算每类平均特征——把新点分给最近类中心。
为什么重要？它解释了极端噪声下鲁棒性的根本来源，颠覆了“过拟合”与“记忆”的泛化叙事。
论文用平均场方法严格证明：无限宽MLP的决策规则等价于nearest-class-mean，且对任意深度、激活函数和噪声类型都成立。有限宽实验完全匹配。
其实它只是很会求平均值——细想又合理。
https://t.co/UrSgC3LQJZ

[恒星 @vintcessun]

以前只能盯着loss曲线和梯度范数看训练状态，其实权重分布本身藏着更细的信号。
每层每个矩阵的幅值分布用双参数Weibull拟合，形状参数k和尺度参数λ就像功能标签加进度条——初始化高斯对应k≈1.2，训练后FFN和输出投影W_o的k锁死在极窄带内，而Q/K投影则偏离Weibull族，偏离程度甚至能反映KV缓存是独立还是共享。
λ随训练单调增长，且与√(η/λ_wd)成线性，直接串联了学习率和权重衰减的联合效应。
https://t.co/xVmuD170Ee

[ueaj @_ueaj]

@DimitrisPapail one of my favorite papers of all time
https://t.co/uClNVZk0W0

[Andrew Gordon Wilson @andrewgwils]

@DimitrisPapail https://t.co/ejNjrSwDDc

[Arthur Gretton @ArthurGretton]

Optimized MMD for Detecting Distribution Shift!

Video:
https://t.co/bTJ0ZkU0zU

and slides
https://t.co/thNvTmjUz9

from the #ICLR26 workshop
https://t.co/WpUCAtJQPO

Covers:
https://t.co/x6syFYarxL
and
https://t.co/OBhBLVwNHc

[Antonio Lupetti @antoniolupetti]

"Advanced Calculus" (Department of Mathematics, Harvard University) is one of the most remarkable advanced mathematics texts I have come across.

The book develops vector spaces, differential calculus, differential equations, integration, differentiable manifolds, differential forms, potential theory, and classical mechanics within a unified mathematical framework.

What I particularly appreciate is the combination of mathematical rigour and geometric intuition. It serves as an excellent introduction to modern analysis and differential geometry.

Definitely worth saving to your bookmarks if you are interested in understanding the mathematical structures underlying advanced calculus.

https://t.co/ZbShWnOpWN

[Kyunghyun Cho @kchonyc]

the full paper is at https://t.co/zhno58PLLl

[Kyunghyun Cho @kchonyc]

@RISignal @soumithchintala @robertnishihara @bschoelkopf @LeonBottou it talks about observational causal inference and how there could be signal in offline data, learning to estimate from purely synthetic data, permutation invariant neural nets, etc. all what i am super excited by and bullish on

[How To Prompt @HowToPrompt__]

Yann Lecun published the most heretical AI paper of the year.

He opens by arguing Magnus Carlsen isn't good at chess and only gets more unhinged from there.

The Turing Award winner and his co-authors dropped a paper demanding the AI industry abandon its biggest obsession, AGI.

Right now, everyone from Silicon Valley CEOs to politicians assumes AGI is the ultimate goal. A machine that can do everything a human can do.

LeCun argues that this entire concept is a biological illusion.

Humans do not possess "general" intelligence. We are highly specialized biological machines, tuned by evolution simply to survive in the physical world.

We only think our intelligence is "general" because we are completely blind to the millions of cognitive tasks we are incapable of comprehending.

Which brings us to the chess argument.

Magnus Carlsen is the greatest human chess player in history. But compared to a modern computer? He is fundamentally terrible.

Our belief that Carlsen is "good" at chess is pure human-centric bias. He isn't objectively good. He's just better than the rest of us, who are biologically awful at it.

LeCun says we need to stop building AI to mimic human generality.

Instead, he proposes a new North Star: SAI.

Superhuman Adaptable Intelligence.

Instead of trying to build a machine that mimics our flawed, biologically-limited brains, we need to embrace extreme specialization.

SAI is about the speed of adaptation.

It is an intelligence that can learn to exceed humans at any specific, economically important task.

More importantly, it is designed to fill the vast skill gaps where humans are fundamentally incapable.

Things like managing global energy grids in real-time. Or predicting complex molecular structures.

The entire AI industry is obsessed with building a digital reflection in our own image.

LeCun's paper is a brutal wake-up call.

[Alex Smola @smolix]

Random projections compress 65,536 dims down to 1,024 — at the price of 67 million multiplications per vector. A lot of work spent in the name of saving work. Ailon & Chazelle's 2009 fix is, literally, a PHD: Φ = P·H·D 🧵

[rohan anil @_arohan_]

Newton Schulz is the fast operator for Matrix Sign the special case when b2=0.0 - this makes Muon fast. This is very cool!

Denman-Beavers coupled newton iteration for inverse square roots and fourth roots is what you need when b2>0.0 - this is what you need to make it really fast.

It’s available in the pytorch package.

[Probability and Statistics @probnstat]

Geometric Deep Learning is an emerging paradigm in machine learning that extends deep learning methods to non-Euclidean domains such as graphs, manifolds, and meshes. Unlike traditional Euclidean-based models, GDL exploits symmetry, invariance, and equivariance principles to design architectures that respect the underlying structure of data. At its core, it generalizes convolution and representation learning beyond regular grids to irregular and structured domains.

In machine learning applications, GDL is widely used through graph neural networks for problems like social network analysis, recommendation systems, and molecular property prediction in chemistry and drug discovery. In deep learning, it enables advances in 3D vision, point cloud processing, and graph-based transformers, while also unifying convolutional neural networks as a special case of geometric operators on lattices. This leads to better inductive biases, improved sample efficiency, and stronger generalization on structured data.

In reinforcement learning, Reinforcement Learning, geometric methods help model multi-agent systems, traffic networks, and relational environments where states and interactions are naturally graph-structured. This supports better coordination, planning, and transfer across agents and environments. Overall, geometric deep learning provides a unifying framework linking ML, DL, and RL by embedding geometry and symmetry into learning systems, enabling more structured, efficient, and generalizable intelligence.

Image: https://t.co/oOnwMycR2b