A Robust Initialization of Residual Blocks for Effective ResNet Training Without Batch Normalization.

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.

A classification framework for exploiting sparse multi-variate temporal features with application to adverse drug event detection in medical records.

BACKGROUND: Adverse drug events (ADEs) as well as other preventable adverse events in the hospital setting incur a yearly monetary cost of approximately $3.5 billion, in the United States alone. Therefore, it is of paramount importance to reduce the impact and prevalence of ADEs within the healthcare sector, not only since it will result in reducing human suffering, but also as a means to substantially reduce economical strains on the healthcare system. One approach to mitigate this problem is to employ predictive models. While existing methods have been focusing on the exploitation of static features, limited attention has been given to temporal features. METHODS: In this paper, we present a novel classification framework for detecting ADEs in complex Electronic health records (EHRs) by exploiting the temporality and sparsity of the underlying features. The proposed framework consists of three phases for transforming sparse and multi-variate time series features into a single-valued feature representation, which can then be used by any classifier. Moreover, we propose and evaluate three different strategies for leveraging feature sparsity by incorporating it into the new representation. RESULTS: A large-scale evaluation on 15 ADE datasets extracted from a real-world EHR system shows that the proposed framework achieves significantly improved predictive performance compared to state-of-the-art. Moreover, our framework can reveal features that are clinically consistent with medical findings on ADE detection. CONCLUSIONS: Our study and experimental findings demonstrate that temporal multi-variate features of variable length and with high sparsity can be effectively utilized to predict ADEs from EHRs. Two key advantages of our framework are that it is method agnostic, i.e., versatile, and of low computational cost, i.e., fast; hence providing an important building block for future exploitation within the domain of machine learning from EHRs.

Lagrangean-Based Combinatorial Optimization for Large-Scale S3VMs.

The process of manually labeling instances, essential to a supervised classifier, can be expensive and time-consuming. In such a scenario the semisupervised approach, which makes the use of unlabeled patterns when building the decision function, is a more appealing choice. Indeed, large amounts of unlabeled samples often can be easily obtained. Many optimization techniques have been developed in the last decade to include the unlabeled patterns in the support vector machines formulation. Two broad strategies are followed: continuous and combinatorial. The approach presented in this paper belongs to the latter family and is especially suitable when a fair estimation of the proportion of positive and negative samples is available. Our method is very simple and requires a very light parameter selection. Several medium- and large-scale experiments on both artificial and real-world data sets have been carried out proving the effectiveness and the efficiency of the proposed algorithm.

Clustering methods for the optimization of atomic cluster structure.

In this paper, we propose a revised global optimization method and apply it to large scale cluster conformation problems. In the 1990s, the so-called clustering methods were considered among the most efficient general purpose global optimization techniques; however, their usage has quickly declined in recent years, mainly due to the inherent difficulties of clustering approaches in large dimensional spaces. Inspired from the machine learning literature, we redesigned clustering methods in order to deal with molecular structures in a reduced feature space. Our aim is to show that by suitably choosing a good set of geometrical features coupled with a very efficient descent method, an effective optimization tool is obtained which is capable of finding, with a very high success rate, all known putative optima for medium size clusters without any prior information, both for Lennard-Jones and Morse potentials. The main result is that, beyond being a reliable approach, the proposed method, based on the idea of starting a computationally expensive deep local search only when it seems worth doing so, is capable of saving a huge amount of searches with respect to an analogous algorithm which does not employ a clustering phase. In this paper, we are not claiming the superiority of the proposed method compared to specific, refined, state-of-the-art procedures, but rather indicating a quite straightforward way to save local searches by means of a clustering scheme working in a reduced variable space, which might prove useful when included in many modern methods.