leAutoencoder

Summary

leAutoencoder is Matteo Peluso’s self-teaching autoencoder prototype. It trains an encoder-decoder through latent consistency between clean and masked reconstruction branches, crop-transformed judging, and SIGReg-style latent regularization instead of direct pixel reconstruction loss.

Role In The Wiki

leAutoencoder belongs to the objective-design branch around JEPA, representation collapse, and decoder-grounded latent learning. It is not a foundation model or peer-reviewed result. Its useful role is as a compact mechanism idea: a decoder can be trained inside a predictive-representation objective when transformed latent agreement keeps outputs tied to the input distribution.

Official Artifacts

• Blog post: https://the-puzzler.github.io/posts/self-teaching-autoencoder/self-teaching-autoencoder.html
• Code: https://github.com/the-puzzler/leautoencoder
• Demo: https://the-puzzler.github.io/posts/self-teaching-autoencoder/latent_brush_demo/

Evidence

• Self-Teaching Autoencoder

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in Self-Teaching Autoencoder. At the entity level, the key transfer question is whether a similar latent-consistency-plus-grounding objective can preserve dense numeric state in time-series models without collapsing into ordinary observation forecasting.

Related Pages

• JEPA
• Representation Collapse
• Self-Supervised Representation Learning
• Latent-Space Predictive Learning
• Vision Foundation Models