Investigating the Limitation of CLIP Models: The Worst-Performing Categories

Contrastive Language-Image Pre-training (CLIP) provides a foundation model by integrating natural language into visual concepts, enabling zero-shot recognition on downstream tasks. It is usually expected that satisfactory overall accuracy can be achieved across numerous domains through well-designed textual prompts. However, we found that their performance in the worst categories is significantly inferior to the overall performance. For example, on ImageNet, there are a total of 10 categories with class-wise accuracy as low as 0\%, even though the overall performance has achieved 64.1\%. This phenomenon reveals the potential risks associated with using CLIP models, particularly in risk-sensitive applications where specific categories hold significant importance. To address this issue, we investigate the alignment between the two modalities in the CLIP model and propose the Class-wise Matching Margin (\cmm) to measure the inference confusion. \cmm\ can effectively identify the worst-performing categories and estimate the potential performance of the candidate prompts. We further query large language models to enrich descriptions of worst-performing categories and build a weighted ensemble to highlight the efficient prompts. Experimental results clearly verify the effectiveness of our proposal, where the accuracy on the worst-10 categories on ImageNet is boosted to 5.2\%, without manual prompt engineering, laborious optimization, or access to labeled validation data.

Parameter-Efficient Long-Tailed Recognition

The "pre-training and fine-tuning" paradigm in addressing long-tailed recognition tasks has sparked significant interest since the emergence of large vision-language models like the contrastive language-image pre-training (CLIP). While previous studies have shown promise in adapting pre-trained models for these tasks, they often undesirably require extensive training epochs or additional training data to maintain good performance. In this paper, we propose PEL, a fine-tuning method that can effectively adapt pre-trained models to long-tailed recognition tasks in fewer than 20 epochs without the need for extra data. We first empirically find that commonly used fine-tuning methods, such as full fine-tuning and classifier fine-tuning, suffer from overfitting, resulting in performance deterioration on tail classes. To mitigate this issue, PEL introduces a small number of task-specific parameters by adopting the design of any existing parameter-efficient fine-tuning method. Additionally, to expedite convergence, PEL presents a novel semantic-aware classifier initialization technique derived from the CLIP textual encoder without adding any computational overhead. Our experimental results on four long-tailed datasets demonstrate that PEL consistently outperforms previous state-of-the-art approaches. The source code is available at https://github.com/shijxcs/PEL.