Hude Liu

We establish the universal approximation capability of single-layer, single-head self- and cross-attention mechanisms