GRAM replaces cosine similarity with the Gramian volume of the parallelotope formed by multiple modality embeddings, and a volume-based contrastive loss improves multimodal retrieval and classification.
Therefore, equation 17 becomes: V ol= p det(G) = p 1 − cos2(θ) = q sin2(θ) = sin(θ)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Gramian Multimodal Representation Learning and Alignment
GRAM replaces cosine similarity with the Gramian volume of the parallelotope formed by multiple modality embeddings, and a volume-based contrastive loss improves multimodal retrieval and classification.