Unverified50% confidenceFactExact time
Feature-level fusion (early fusion) aggregates features from multiple sources after feature extraction and before similarity computation, preserving more detail but being sensitive to noise; score-level fusion (late fusion) first computes similarities independently and then merges them, being more robust but potentially losing cross-view association information.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCLIP 用对比学习（Contrastive Learning）替代了此前的监督分类范式，通过最大化匹配图文对相似度、最小化不匹配对相似度进行自监督学习63% similarUnverifiedMean aggregation is more robust to variations in row count, while max is more sensitive to prominent features; the two can be concatenated and used together.63% similarUnverifiedLate fusion vision performs worse than native multimodal training (early fusion) on visual aesthetics and design generation tasks.61% similarUnverifiedDiffusion features are extracted by performing several denoising steps on a noisy image during inference and then intercepting the activations of intermediate layers in the backbone network as visual representations for downstream discriminative tasks.60% similarUnverifiedSCARCE's theoretical connection reveals the equivalence between adversarial confusion and feature reweighting, offering a new perspective for understanding label-free generalization.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115463API
curl https://kongchang.com/api/v1/knowledge/claims/1115463MCP
get_claim(id=1115463)