Noise Stability of Transformer Models

Themistoklis Haris, Zihan Zhang, Yuichi Yoshida

Understanding simplicity biases in deep learning offers a promising path toward developing reliable AI. A common metric for this, inspired by Boolean function analysis, is average…