RESUMO
Batch normalization is an essential component of all state-of-the-art neural networks architectures. However, since it introduces many practical issues, much recent research has been devoted to designing normalization-free architectures. In this brief, we show that weights initialization is key to train ResNet-like normalization-free networks. In particular, we propose a slight modification to the summation operation of a block output to the skip-connection branch, so that the whole network is correctly initialized. We show that this modified architecture achieves competitive results on CIFAR-10, CIFAR-100 and ImageNet without further regularization nor algorithmic modifications.
RESUMO
OBJECTIVES: We evaluated a monocentric SLE cohort in order to assess the frequency of Lupus comprehensive disease control (LupusCDC), a condition defined by the achievement of remission and the absence of damage progression. METHODS: Our longitudinal analysis included SLE patients with 5-years follow-up and at least one visit per year. Disease activity was assessed by SLE Disease Activity Index 2000 (SLEDAI-2K) and three different remission levels were evaluated (Complete Remission, CR; Clinical remission off-corticosteroids; clinical remission on-corticosteroids). Chronic damage was assessed according to SLICC Damage Index (SDI). LupusCDC was defined as remission achievement for at least one year plus absence of chronic damage progression in the previous one year. A machine learning based analysis was carried out, applying and comparing Nonlinear Support Vector Machines (SVM) models and Decision Trees (DT), whereas features ranking was performed with the ReliefF algorithm. RESULTS: We evaluated 172 patients [M/F 16/156, median age 49 years (IQR 16.7), median disease duration 180 months (IQR 156)]. SDI values (baseline mean±SD 0.7 ± 1.1) significantly increased during the follow-up period. In all time-points analyzed, LupusCDC including CR was the most frequently detected. The failure to reach this condition was significantly associated with renal involvement and with the intake of immunosuppressant drugs and glucocorticoid (GC). Ten patients (5.8%) have maintained LupusCDC during the whole 5-year follow-up: these patients had never presented renal involvement and showed lower prevalence of anti-phospholipid antibodies (p = 0.0001). Finally, the prevalence of GC intake was significantly lower (p = 0.0001). The application of machine learning models showed that the available features were able to provide significant information to build predictive models with an AUC score of 0.703 ± 0.02 for DT and 0.713 ± 0.02 for SVM. CONCLUSIONS: Our data on a monocentric cohort suggest that the LupusCDC can efficaciously merge into one outcome SLE-related disease activity and chronic damage in order to perform an all-around evaluation of SLE patients.