Concetti Chiave
Combining deep metric learning and synthetic data generation using diffusion models to improve out-of-distribution detection in classification models.
Sintesi
The paper presents a novel approach for out-of-distribution (OOD) detection in classification models by combining deep metric learning and synthetic data generation using diffusion models.
The key highlights are:
Synthetic OOD data generation using a label-mixup approach with Denoising Diffusion Probabilistic Models (DDPMs). This generates diverse and meaningful OOD samples by interpolating the one-hot encodings of different classes.
Exploration of recent advancements in metric learning, such as SphereFace, CosFace, ArcFace, and AdaCos, as alternative loss functions for training OOD detectors. These metric learning-based loss functions aim to increase inter-class variation and reduce intra-class variation in the feature space.
Evaluation of the proposed method against well-known OOD detection approaches. The results show that the combination of synthetic outlier exposure and metric learning-based loss functions outperforms the baseline methods in conventional OOD detection metrics (AUROC and AUPR).
The proposed outlier exposure approach leads to a significant performance improvement across various loss functions, including the vanilla softmax and metric learning-based losses, compared to the baseline models without access to training OOD data.
The in-distribution classification accuracy of the models trained with the proposed outlier exposure remains largely unchanged, demonstrating the ability to detect OOD samples without compromising closed-set performance.
Statistiche
The maximum cosine similarity between the normalized features and weights is used as the OOD score function.
The threshold for OOD detection is computed from a validation set at 95% true positive rate (TPR).
Citazioni
"By combining deep metric learning and synthetic data generation using diffusion models, our proposed method offers a promising solution for improving OOD detection."
"Our results show that our approach outperform baseline methods in conventional OOD detection metrics (AUROC and AUPR)."