This article presents findings on the benefits of L2 normalization in neural networks, particularly in improving out-of-distribution (OoD) detection. The study indicates that utilizing L2 normalization can lead to competitive results in less time, achieving notable improvements in AUROC scores for ResNet models. The results reveal a promising reduction in training time, showing that superior OoD detection can be achieved without extensive computational resources, emphasizing the potential for advancing deep learning techniques.
In our experiments, we find that L2 normalization over feature space not only reduces training time significantly but also improves out-of-distribution (OoD) detection performance.
For ResNet18, we achieved higher mean AUROC scores compared to the baseline in just 60 epochs, showcasing the efficiency of L2 normalization.
Collection
[
|
...
]