Unverified50% confidenceFactExact time
When using contrastive learning objectives, the damage of noisy labels to vision-language alignment is relatively controllable; generative objectives (such as image captioning) are more sensitive to label errors.
1
Sources
50%
Confidence
Long-term
Relevance
11/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn vision-language pretraining, soft labels or label smoothing can be applied to the image-text contrastive loss, replacing hard labels with a mixture of a uniform distribution and the true label, thereby preventing the model from being overconfident about a single incorrect label.78% similarUnverifiedAdversarial sample attacks can add noise to images that is nearly imperceptible to the human eye yet causes models to produce incorrect classifications69% similarUnverifiedImage moderation faces adversarial evasion problems, such as bypassing detection by slightly modifying pixels or adding noise.66% similarUnverifiedUnder noise interference, high-level semantic features are damaged first, while low-level features can still be extracted; humans can proactively shift attention from the semantic level to the perceptual level, which is a manifestation of cognitive flexibility in coping with visual uncertainty.66% similarUnverified开发者指出abliterate一个模型并不会让图像输出更加去审查，编码器真正价值在于提示词增强与视觉解码场景66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1146926API
curl https://kongchang.com/api/v1/knowledge/claims/1146926MCP
get_claim(id=1146926)