Mostafa Dehghani

Dehghani co-developed the Universal Transformer, which repeatedly refines its internal representation instead of using a fixed sequence of layers. He later helped scale vision Transformers to tens of billions of parameters, testing how far general language-model ideas could transfer to images.