COMBINING CONVOLUTIONS AND SELF-ATTENTIONS FOR EFFICIENT VISION MODELS

Summary from the TLDR AI newsletter

GLMix is a new method for integrating convolutions and multi-head self-attentions (MHSAs) at different granularity levels in vision tasks. Convolutions handle fine-grained local features, while MHSAs operate on coarse-grained semantic slots for global insights.