Back to AI Lab
ArXiv Paper

Deep Delta Learning

Yifan Zhang, Yifeng Liu, Mengdi Wang +1January 1, 2026

Summary

The authors replace standard residual skip connections with a learnable "Delta" operator that can flexibly distort the identity path. This lets deep nets control how much old information to erase versus new information to write, improving how they model complex dynamics while keeping training stable.

Related Content