News
Generalizable Vision-Language Few-Shot Adaptation with Predictive Prompts and Negative Learning
A new few-shot vision-language method claims a 4.61% average gain at 16-shot by only suppressing the classes a query is actually likely to confuse. For game teams using CLIP-like models …
Sriram Mandalika
·
vision-language models
few-shot learning
prompt engineering
contrastive learning
robustness
+1