Neural networks are notorious for being overconfident predictors, posing a significant challenge to their safe deployment in real-world applications. While feature normalization has garnered considerable attention within the deep learning literature, current train-time regularization methods for Out-of-Distribution(OOD) detection are yet to fully exploit this potential. Indeed, the naive incorporation of feature normalization within neural networks does not guarantee substantial improvement in OOD detection performance. In this work, we introduce T2FNorm, a novel approach to transforming features to hyperspherical space during training, while employing non-transformed space for OOD-scoring purposes. This method yields a surprising enhancement in OOD detection capabilities without compromising model accuracy in in-distribution(ID). Our investigation demonstrates that the proposed technique substantially diminishes the norm of the features of all samples, more so in the case of out-of-distribution samples, thereby addressing the prevalent concern of overconfidence in neural networks. The proposed method also significantly improves various post-hoc OOD detection methods.

翻译：神经网络因倾向于给出过度自信的预测而著称，这对其在现实世界应用中的安全部署构成了重大挑战。尽管特征归一化已在深度学习文献中引起广泛关注，但当前用于分布外（OOD）检测的训时正则化方法尚未充分利用这一潜力。事实上，在神经网络中简单加入特征归一化并不能保证OOD检测性能的显著提升。在本工作中，我们提出T2FNorm，一种新颖的方法：在训练期间将特征变换到超球面空间，同时采用非变换空间进行OOD评分。该方法在不降低模型在分布内（ID）准确率的前提下，意外地增强了OOD检测能力。我们的研究表明，所提技术显著降低了所有样本的特征范数，其中对OOD样本的影响尤为明显，从而缓解了神经网络中普遍存在的过度自信问题。此外，该方法还能显著提升多种事后OOD检测方法的性能。