Skip to content

Comment on State-space models can learn in-context by gradient descent

Comments

We show that SSMs with local self-attention, a form of input-dependent input processing, can perform in-context learning analogously to transformers, i.e. through gradient descent steps on an implicit linear regression problem.

I don't understand. The benefit of SSMs is better scalability than self-attention. Now this adds self-attention back?

It adds a very local sliding window attention, the context is only 3 adjacent frames per step. They need the access to adjacent frames to show the implicit model gradient computation but I didn't yet follow the derivation for why this is so.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.