How Does Fine-Tuning Impact Out-of-Distribution Detection for Vision-Language Models?

Abstract

Recent large vision-language models such as CLIP have shown remarkable out-of-distribution (OOD) detection and generalization performance. However, their zero-shot in-distribution (ID) accuracy is often limited for downstream datasets. Recent CLIP-based fine-tuning methods such as prompt learning have demonstrated significant improvements in ID classification and OOD generalization where OOD labels are available. Nonetheless, it remains unclear whether the model is reliable to semantic shifts without OOD labels. In this paper, we aim to bridge the gap and present a comprehensive study to understand how fine-tuning impact OOD detection for few-shot downstream tasks. By framing OOD detection as multi-modal concept matching, we establish a connection between fine-tuning methods and various OOD scores. Our results suggest that a proper choice of OOD scores is essential for CLIP-based fine-tuning. In particular, the maximum concept matching (MCM) score provides a promising solution consistently. We also show that prompt learning demonstrates the state-of-the-art OOD detection performance over the zero-shot counterpart.

Appendices

Appendix A Dataset Details

1.1 Details on ID and OOD Dataset Construction

For ID datasets, we follow the same construction as in previous works (Zhang et al., 2022; Zhou et al., 2022a, b). Detailed instructions on dataset installation can be found in https://github.com/KaiyangZhou/CoOp/blob/main/DATASETS.md. For OOD datasets, Huang and Li (2021) curate a collection of subsets from iNaturalist Van Horn et al. (2018), SUN Xiao et al. (2010), Places Zhou et al. (2017), and Texture Cimpoi et al. (2014) as large-scale OOD datasets for ImageNet-1k, where the classes of the test sets do not overlap with ImageNet-1k. Detailed instructions can be found in https://github.com/deeplearning-wisc/large_scale_ood.

Appendix B: Additional Results

1.1 B.1 ID Accuracy

While we primarily focus on the OOD detection performance of CLIP-based fine-tuning methods, we present the results of the ID accuracy for each dataset based on CLIP-B/16 in Table 6 for completeness. Further results on the ID accuracy with various datasets and architectures can be seen in Zhou et al. (2022a), Zhou et al. (2022b), and  Zhang et al. (2022).

1.2 B.2 OOD detection performance based on visual features alone

In this section, we explore several commonly used OOD detection scores solely based on the visual branch of CLIP models. Specifically, we consider the Mahalanobis score (Lee et al., 2018) on the penultimate layer of the visual encoder and MSP (Hendrycks & Gimpel, 2017), Energy (Liu et al., 2020), and KL Matching (Hendrycks et al., 2022) scores on the logit layer after linear probing the visual encoder. The results are summarized in Table 7, .based on 16-shot Caltech-101 (ID). We can see that the Mahalanobis score does not yield promising performance because 1) the feature embeddings from the visual encoder of CLIP may not follow class-conditional Gaussian distributions, 2) it is challenging to estimate the mean and especially covariance matrix when the number of samples is much smaller than the feature dimension in the few-shot setting. On the other hand, the OOD scores based on fine-tuned logit layer result in worse performance compared to the MCM score. One major reason is that fine-tuning CLIP in the few-shot setting is prone to overfitting the downstream ID dataset, making the model less reliable. This further highlights the importance of choosing OOD detection scores fitted to parameter-efficient fine-tuning methods.

Table 8 OOD detection performance on two OOD additional test sets for ImageNet-1k (ID). We train CLIP-B/16 with CoOp

Table 9 OOD detection performance based on $S_{\text {MSP}}$ score. The average performance for most adaptation methods is much worse than using $S_{\text {MS}}$ (Table 1) and $S_{\text {MCM}}$ (Table 3)

1.3 B.3 Additional Results on ImageNet-1k

In this section, we consider two additional OOD test sets ImageNet-O (Hendrycks et al., 2021) and OpenImage-O (Wang et al., 2022) for ImageNet-1k (ID). OpenImage-O is a subset curated from the test set of OpenImage-V3 (Krasin et al., 2017) containing a diverse set of categories. ImageNet-O is a challenging OOD dataset that contains naturally adversarial examples for ImageNet-1k. The results are shown in Table 8. The model (CLIP-B/16) is trained with CoOp. We can see that: 1) The performance on ImageNet-O is generally worse than the rest of OOD test sets (iNaturalist, Textures, SUN, Places) in Sect. 4.3, suggesting that this task remains challenging in the context of few-shot prompt learning. 2) MCM score still performs the best compared to MS and MSP on both OOD test sets, consistent with our previous observations, which further highlights the importance of softmax and temperature scaling for OOD detection with fine-tuning.

Table 10 Comparison with additional OOD scores on Caltech-101 (ID). $S_{\text {KL}}$ stands for the KL matching score (Hendrycks et al., 2022) and $S_{\text {Energy}}$ denotes the energy score (Liu et al., 2020)

1.4 B.4 Alternative OOD Scores

In this section, we investigate the performance with several alternative OOD scoring functions based on the cosine similarities of input $\varvec{x}$ with the k-th label $s_k(\varvec{x})$, $k\in \{1,2,...,K\}$ (defined in Sect. 3.2). Specifically, we consider the energy and the KL matching score for each adaptation method and summarize the results based on Caltech-101 (ID) in Table 10. We observe that 1) using the energy score, all adaptation methods significantly enhance the performance over the zero-shot baseline (ZOCLIP). 2) the general performance vastly improves when utilizing the KL Matching score. However, even the highest achieved performance (FPR95 at 7.91 with CoCoOp) falls short when compared to the MCM score (FPR95 at 5.02 with CoCoOp).