IC-124Direct comparison of CLIP image embeddings with CLAP audio embeddings achieves near-chance retrieval, while logsumexp bridging through the shared language modality recovers 62% recall@10 on AudioSet

Yongwei Che, Benjamin Eysenbach

SourceThe "Law'' of the Unconscious Contrastive Learner: Probabilistic Alignment of Unpaired Modalities

The paper evaluates whether the image encoder of CLIP and the audio encoder of CLAP can be directly compared via normalized dot product for audio-visual retrieval. Direct comparison yields only 14% recall@10 on AudioSet, barely above the 10% chance level for a 10-way task. Applying the logsumexp (monte carlo) method, which marginalizes over the intermediate language modality using the AudioSet ontology as the sample space, raises performance to 62% recall@10. This demonstrates that two independently trained contrastive models cannot be naively plugged together without an explicit bridging procedure.

Evidence
correlational
Key metric
14% recall@10 (direct CLIP image vs CLAP audio) vs 62% recall@10 (logsumexp via language) on AudioSet
Caveat
The logsumexp method uses only the AudioSet ontology (600-700 language descriptions) as the intermediate sample space, which limits the approximation quality.
Model
CLIP / CLIP-ViT (LC), CLAP
Concepts
Failure mode
Datasets
AudioSet [eval]
Related work
LanguageBind [compared-to]
Related findings
IC-125, IC-126, IC-127
Extraction
automatic-extraction