Maniparambil et al.

Showed well-trained vision encoders exhibit high semantic similarity with language encoders regardless of training paradigm