They swapped dot product for a psych-based asymmetric Tversky similarity, making nets more fire and less basic, no cap.
New Tversky layers flex on image and language tasks—ResNet-50 got an 8.9% boost and GPT-2 cut perplexity by 7.5%, big rizz.
It’s hella interpretable: prototypes show real features so you get the tea on model decisions, main character energy.
Get notified when new stories are published for "Gen-Z Tech News"