Ruijin Hua

[Ruijin Hua @Helios_Hua]

@Zhaoj6317 No, it doesn’t require specially collected data. All we need to do is find suitable real-world conditions where our Split JEPA can fit. I’m glad you like our work!

[Ruijin Hua @Helios_Hua]

A latent representation may capture the entire state of a world without revealing its underlying structure.

Our new work SplitJEPA, explores how to identify the invariant-variant structure without reconstructing observations.

https://t.co/5E3FriBd48

[Ruijin Hua @Helios_Hua]

We hope this work advances understanding of reconstruction-free latent identification.

📄 Paper: https://t.co/q2U5pfDRk1
💻 Code: https://t.co/KStMFny1iw

Grateful to have worked with @YujiaZheng9 @zhuokaiz @c93l6IhoSgV2Iqi on this project!

[Ruijin Hua @Helios_Hua]

Why does this matter in the real world?

Real-world observations mix meaningful structure with changing conditions and redundant details. SplitJEPA separates invariant and variant factors without reconstruction, potentially reducing computational overhead and unnecessary details.