Beyond Accuracy: On the Effects of Fine-tuning Towards Vision-Language Model's Prediction Rationality

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vision-Language Models (VLMs) like CLIP are widely used.
• Fine-tuning VLMs is common in safety-critical domains.
• Prediction rationality (correctness and valid evidence) is vital in these domains.

Purpose of the Study:

• Investigate the impact of fine-tuning on VLM prediction rationality.
• Introduce novel metrics: Prediction Trustworthiness and Inference Reliability.

Main Methods:

• Conducted extensive experiments across various settings.
• Evaluated fine-tuned VLMs using the proposed metrics.
• Assessed model performance under distributional shifts.

Main Results:

• Fine-tuning improved prediction accuracy but sometimes relied on invalid evidence.
• Fine-tuned VLMs showed higher accuracy when using valid evidence.
• Findings remained consistent across different settings and shifts.

Conclusions:

• Standard fine-tuning may decrease VLM trustworthiness by increasing reliance on invalid evidence.
• Valid evidence identification is key for reliable predictions from fine-tuned VLMs.
• Research offers new insights into VLM fine-tuning for critical applications.