Enhancing out-of-distribution learning in computer vision through dominant feature masking

Enhancing out-of-distribution learning in computer vision through dominant feature masking

연구 분야: Artificial Intelligence

학회: Pattern Analysis and Applications

Out-of-distribution (OOD) learning presents a major challenge in machine learning as models must effectively generalize to previously unseen data. This challenge is prevalent in deep learning models, which tend to focus on the most dominant features in images. This narrow focus impedes OOD learning, where critical features are concealed or absent during testing, leading to reduced prediction accuracy. To address this issue, we introduce a novel data augmentation approach termed Dominant Feature Masking (DFM), inspired by human visual holistic processing. DFM strategically conceals and reveals the most prominent features within images, allowing neural networks to simultaneously capture both dominant and non-dominant attributes, thereby enhancing adaptability to OOD data. We evaluated DFM using a novel set of learning challenges termed Versatile Evaluation Benchmark (VEB), which assesses model performance on three distinct tasks: (i) augmented MNIST images to test resilience against diverse transformations; (ii) a novel dataset of unseen image classes to examine performance on new instances within familiar categories; and (iii) a dataset created by DALL-E to challenge class differentiation with artificially mixed features. Our results demonstrate that DFM significantly improves OOD generalization compared to traditional augmentation techniques, achieving marked enhancements across various conditions without compromising in-distribution testing accuracy. These findings underscore the potential of DFM to improve the performance of computer vision systems in various real-world scenarios, making them more robust and adaptable to unexpected data variations. By leveraging VEB, researchers will gain a deeper understanding of their models’ generalization performance, ensuring that CNNs are well-equipped to handle the complexities of real-world applications. The source code and VEB datasets are available at https://github.com/DeepVisionary/DFM.