Oliver Sieberling presents dynamic short convolutions for Transformers

Sieberling's talk introduces dynamic short convolutions, a scalable primitive using input-dependent filters to adaptively aggregate local context in Transformer language models.