Interpolated Adversarial Training: Achieving robust neural networks without sacrificing too much accuracy.

Adversarial robustness has become a central goal in deep learning, both in the theory and the practice. However, successful methods to improve the adversarial robustness (such as adversarial training) greatly hurt generalization performance on the unperturbed data. This could have a major impact on how the adversarial robustness affects real world systems (i.e. many may opt to forego robustness if it can improve accuracy on the unperturbed data). We propose Interpolated Adversarial Training, which employs recently proposed interpolation based training methods in the framework of adversarial training. On CIFAR-10, adversarial training increases the standard test error ( when there is no adversary) from 4.43% to 12.32%, whereas with our Interpolated adversarial training we retain the adversarial robustness while achieving a standard test error of only 6.45%. With our technique, the relative increase in the standard error for the robust model is reduced from 178.1% to just 45.5%. Moreover, we provide mathematical analysis of Interpolated Adversarial Training to confirm its efficiencies and demonstrate its advantages in terms of robustness and generalization.

Interpolation consistency training for semi-supervised learning.

We introduce Interpolation Consistency Training (ICT), a simple and computation efficient algorithm for training Deep Neural Networks in the semi-supervised learning paradigm. ICT encourages the prediction at an interpolation of unlabeled points to be consistent with the interpolation of the predictions at those points. In classification problems, ICT moves the decision boundary to low-density regions of the data distribution. Our experiments show that ICT achieves state-of-the-art performance when applied to standard neural network architectures on the CIFAR-10 and SVHN benchmark datasets. Our theoretical analysis shows that ICT corresponds to a certain type of data-adaptive regularization with unlabeled points which reduces overfitting to labeled points under high confidence values.

Rhabdomyolysis in obese trauma patients.

Patients sustaining traumatic injuries are at risk for development of rhabdomyolysis. The effect of obesity on this risk is unknown. This study attempted to characterize the role of obesity in the development of rhabdomyolysis after trauma. This was a retrospective review of all trauma patients with creatine kinase (CK) levels admitted to the surgical intensive care unit (SICU) at a Level I trauma center from February 2011 until July 2013. Patients were divided based on their body mass index (BMI): overweight/obese group with BMI 25 kg/m(2) or greater and nonoverweight/obese group with BMI less than 25 kg/m(2). Primary outcome was CK greater than 10,000 U/L. During the 30-month study period, 198 trauma patients with available CK levels were admitted to the SICU. The majority (27.8%) of patients were involved in a motor vehicle collision. There were 96 patients (48.4%) with BMI 25 kg/m(2) or greater and 102 (51.5%) with BMI less than 25 kg/m(2). There was no difference in creatinine levels between the two groups (1.5 ± 1.2 mg/dL vs 1.5 ± 1.4 mg/dL, P = 0.83). BMI 25 kg/m(2) or greater was independently associated with the development of CK greater than 10,000 U/L (14.6 vs 4.9%; adjusted odds ratio, 3.03; P = 0.04). Patients with BMI 25 kg/m(2) or greater are at a significantly higher risk for rhabdomyolysis after trauma. Aggressive CK level monitoring to prevent rhabdomyolysis in this population is strongly encouraged.

Making sense of information in noisy networks: human communication, gossip, and distortion.

Information from others can be unreliable. Humans nevertheless act on such information, including gossip, to make various social calculations, thus raising the question of whether individuals can sort through social information to identify what is, in fact, true. Inspired by empirical literature on people's decision-making when considering gossip, we built an agent-based simulation model to examine how well simple decision rules could make sense of information as it propagated through a network. Our simulations revealed that a minimalistic decision-rule 'Bit-wise mode' - which compared information from multiple sources and then sought a consensus majority for each component bit within the message - was consistently the most successful at converging upon the truth. This decision rule attained high relative fitness even in maximally noisy networks, composed entirely of nodes that distorted the message. The rule was also superior to other decision rules regardless of its frequency in the population. Simulations carried out with variable agent memory constraints, different numbers of observers who initiated information propagation, and a variety of network types suggested that the single most important factor in making sense of information was the number of independent sources that agents could consult. Broadly, our model suggests that despite the distortion information is subject to in the real world, it is nevertheless possible to make sense of it based on simple Darwinian computations that integrate multiple sources.