- Bayesian Various-Results Vector Autoregressive Fashions for Inference of Mind Connectivity Networks and Covariate Results in Pediatric Traumatic Mind Harm
Authors: Yangfan Ren, Nathan Osborne, Christine B. Peterson, Dana M. DeMaster, Linda Ewing-Cobbs, Marina Vannucci
Summary: On this paper, we develop an analytical method for estimating mind connectivity networks that accounts for topic heterogeneity. Extra particularly, we take into account a novel extension of a multi-subject Bayesian vector autoregressive mannequin that estimates group-specific directed mind connectivity networks and accounts for the consequences of covariates on the community edges. We undertake a versatile method, permitting for (probably) non-linear results of the covariates on edge energy through a novel Bayesian nonparametric prior that employs a weighted combination of Gaussian processes. For posterior inference, we obtain computational scalability by implementing a variational Bayes scheme. Our method permits simultaneous estimation of group-specific networks and collection of related covariate results. We present improved efficiency over competing two-stage approaches on simulated information. We apply our methodology on resting-state fMRI information from kids with a historical past of traumatic mind harm and wholesome controls to estimate the consequences of age and intercourse on the group-level connectivities. Our outcomes spotlight variations within the distribution of dad or mum nodes. In addition they recommend alteration within the relation of age, with peak edge energy in kids with traumatic mind harm (TBI), and variations in efficient connectivity energy between women and men.
2. Attacking Bayes: On the Adversarial Robustness of Bayesian Neural Networks
Authors: Yunzhen Feng, Tim G. J. Rudner, Nikolaos Tsilivis, Julia Kempe
Summary: Adversarial examples have been proven to trigger neural networks to fail on a variety of imaginative and prescient and language duties, however current work has claimed that Bayesian neural networks (BNNs) are inherently strong to adversarial perturbations. On this work, we study this declare. To review the adversarial robustness of BNNs, we examine whether or not it’s doable to efficiently break state-of-the-art BNN inference strategies and prediction pipelines utilizing even comparatively unsophisticated assaults for 3 duties: (1) label prediction below the posterior predictive imply, (2) adversarial instance detection with Bayesian predictive uncertainty, and (3) semantic shift detection. We discover that BNNs skilled with state-of-the-art approximate inference strategies, and even BNNs skilled with Hamiltonian Monte Carlo, are extremely prone to adversarial assaults. We additionally determine varied conceptual and experimental errors in earlier works that claimed inherent adversarial robustness of BNNs and conclusively exhibit that BNNs and uncertainty-aware Bayesian prediction pipelines will not be inherently strong in opposition to adversarial assaults.