Vanilla training canbeconsidered asaspecial case where no perturbation is allowed, i.e., zero adversarial budget. Therefore, we focus on the impact of the adversarial budget size on the loss landscape.

Perception, in theoretical neuroscience, has been modeled as the encoding of external stimuli into internal signals, which are then decoded.

Existing image-text modality alignment in Vision Language Models (VLMs) treats each text token equally in an autoregressive manner.

Clearly, due to nonconvexity, the task of finding a minimizer of (1) canbequite challenging andpotentially limits what canbeguaranteed theoretically about solving problems with form(1).