We present a theoretical analysis of CLIP, showing how temperature optimization enables adaptation to the intrinsic dimension of shared features in multi-modal data (poster).
A more recent work proposes IndiSeek which learns modality-specific features that are independent of shared features.