Abstract:Out-of-distribution (OOD) detection is critical to ensure the safe deployment of deep learning models in critical applications. Deep learning models can often misidentify OOD samples as in-distribution (ID) samples. This vulnerability worsens in the presence of spurious correlation in the training set. Likewise, in fine-grained classification settings, detection of fine-grained OOD samples becomes inherently challenging due to their high similarity to ID samples. However, current research on OOD detection has largely ignored these challenging scenarios, focusing instead on relatively easier (conventional) cases. In this work, we present a unified Approach to Spurious, fine-grained, and Conventional OOD Detection (ASCOOD). First, we propose synthesizing virtual outliers from ID data by approximating the destruction of invariant features. We identify invariant features with the pixel attribution method using the model being learned. This approach eliminates the burden of curating external OOD datasets. Then, we simultaneously incentivize ID classification and predictive uncertainty towards the virtual outliers leveraging standardized feature representation. Our approach effectively mitigates the impact of spurious correlations and encourages capturing fine-grained attributes. Extensive experiments across six datasets demonstrate the merit of ASCOOD in spurious, fine-grained, and conventional settings. The code is available at: <a class="link-external link-https" href="https://github.com/sudarshanregmi/ASCOOD/" rel="external noopener nofollow">this https URL</a>
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **In deep - learning models, how to effectively detect out - of - distribution (OOD) samples, especially in the presence of spurious correlations and fine - grained classification**.
Specifically, deep - learning models may misidentify OOD samples as in - distribution (ID) samples when encountering them. This misidentification problem is particularly serious in the following two situations:
1. **Spurious Correlation**: When there are spurious correlations in the training data, the model may rely on these unstable features for prediction. For example, in an image classification task, the model may rely on the background or artifacts in the image to classify, rather than truly focusing on the features of the target object. This will lead to the model making wrong high - confidence predictions for OOD samples during testing.
2. **Fine - grained Classification**: In fine - grained classification tasks, the differences between OOD samples and ID samples are very subtle, making detection difficult. For example, in species classification, new species may be very similar to known species but still belong to OOD samples.
Most current OOD detection methods mainly focus on relatively simple traditional scenarios and ignore the above - mentioned challenging scenarios. Therefore, this paper proposes a new method - **ASCOOD (Approach to Spurious, fine - grained, and Conventional OOD Detection)** to solve these problems.
### Main contributions of ASCOOD:
1. **Virtual outlier synthesis**: Synthesize virtual outliers by perturbing invariant features while keeping environmental features almost unchanged. This method does not require an external OOD dataset.
2. **Joint training framework**: Design a joint training objective that not only optimizes the accuracy of ID classification but also improves the uncertainty prediction of synthesized virtual outliers.
3. **Standardized feature representation**: Use standardized feature representation to constrain the over - fitting behavior of the model, thereby improving the performance of OOD detection.
### Experimental results:
- On benchmark datasets such as Waterbirds and CelebA, ASCOOD significantly outperforms 28 existing methods, especially performing well on the FPR@95 metric.
- In traditional OOD detection tasks such as CIFAR - 10/100, ASCOOD also achieves significant performance improvement.
- In fine - grained classification tasks (such as Aircraft and Car datasets), ASCOOD also performs excellently, surpassing several existing methods.
In conclusion, this paper effectively solves the problems of spurious correlations and fine - grained classification in OOD detection by introducing virtual outlier synthesis and a joint training framework, improving the security and reliability of the model.