Recent remarkable success in the deep-learning industries has unprecedentedly
increased the need for reliable model deployment. For example, the model should
alert the user if the produced model outputs might not be reliable. Previous
studies have proposed various methods to solve the Out-of-Distribution (OOD)
detection problem, however, they generally require a burden of resources. In
this work, we propose a novel and simple method, Multiple Input Mixup (MIM).
Our method can help improve the OOD detection performance with only single
epoch fine-tuning. Our method does not require training the model from scratch
and can be attached to the classifier simply. Despite its simplicity, our MIM
shows competitive performance. Our method can be suitable for various
environments because our method only utilizes the In-Distribution (ID) samples
to generate the synthesized OOD data. With extensive experiments with CIFAR10
and CIFAR100 benchmarks that have been largely adopted in out-of-distribution
detection fields, we have demonstrated our MIM shows comprehensively superior
performance compared to the SOTA method. Especially, our method does not need
additional computation on the feature vectors compared to the previous studies.
All source codes are publicly available at
this https URL

我们提出了一种新颖且简单的方法，称为 Multiple Input Mixup（MIM），它可以通过单次 fine-tuning 改善 OOD 检测性能，而无需从头训练模型。我们的方法仅利用 ID 样本生成合成的 OOD 数据，且在 CIFAR10 和 CIFAR100 基准测试中展示出与 SOTA 方法相比更出色的性能，且不需要对特征向量进行额外计算。