Julia Cohen

Transformers no longer require specialized silicon to run in orbit: reformulating self-attention into gated convolutional projections enables robust, noise-resilient visual attention directly on off-the-shelf CNN edge accelerators.