Kernkonzepte
Data balancing in multimodal systems can mitigate biases but may have mixed impacts on model quality.
Statistiken
CLIP models can absorb societal stereotypes.
Fine-tuning on balanced data counters representation biases effectively.
Data balancing can improve classification but hurt retrieval performance.
Zitate
"Data balancing has a mixed impact on quality: it tends to improve classification but can hurt retrieval."
"Fine-tuning is effective in countering representation biases, though its impact diminishes for association biases."