Twitxy Saved

Yilun Kuang

@KuangYilun

NYC yilunkuang.github.io Joined March 2019
518 Following 535 Followers 133 Posts

Latest posts 5

Yilun Kuang @KuangYilun
Check out this writeup about our Rectified LpJEPA paper from the NYU CDS team!
CDS PhD student @KuangYilun, CDS founding director @ylecun, former CDS Faculty Fellow @timrudner, and others successfully applied biological sparsity to AI.

Their new technique allows computer vision models to ignore 90% of data without losing accuracy.

nyudatascience.medium.com/new-representa…
View quoted post on X
0 replies 5 reposts 26 likes 4.2K views View on X
Yilun Kuang @KuangYilun
Very cool work in identifying why diffusion struggles in representation space!

From the representation learning side, we have also developed a technique in our Radial-VCReg paper (arxiv.org/abs/2602.14272) to prevent radial entropy collapse by regularizing the entropy of the radial distributions of neural network features.

The radial distribution is one-dimensional, and hence entropy estimations using m-spacing estimators are quite tractable and easy to optimize.

Would be interesting to explore whether encouraging non-degenerate radial distributions at the encoder level could also mitigate the geometric interference observed during training.

This is joint work with @yashdagad, Deep Chakraborty, Erik Learned-Miller, @randall_balestr, @timrudner, and @ylecun
arxiv.org Radial-VCReg: More Informative Representation Learning Through... Self-supervised learning aims to learn maximally informative representations, but explicit information maximization is hindered by the curse of dimensionality. Existing methods like VCReg address...
Amandeep Kumar @Amandeep__kumar Feb 11
🚀 Unlocking Standard Diffusion Transformers on Representation Encoders

Why do standard DiTs fail to converge on high-dimensional features like DINOv2? 📉 We found the answer isn't just "more parameters"—it's Geometry.

Introducing Riemannian Flow Matching with Jacobi Regularization (RJF)

📄 Paper: arxiv.org/abs/2602.10099
View quoted post on X
2 replies 29 reposts 186 likes 16K views View on X
Yilun Kuang @KuangYilun
How do we build sparsity into JEPA representations by design, while preserving task-relevant information?

Introducing Rectified LpJEPA, a JEPA architecture that learns sparse, non-negative, informative representations through principled distributional regularization. 📐

📄 Paper: arxiv.org/abs/2602.01456
💻 Code: github.com/YilunKuang/rec…
📝 Blog: yilunkuang.github.io/blog/2026/rect…

(1/n)