skip to main content

Title: Dangers of Bayesian Model Averaging under Covariate Shift
Approximate Bayesian inference for neural networks is considered a robust alternative to standard training, often providing good performance on out-of-distribution data. However, Bayesian neural networks (BNNs) with high-fidelity approximate inference via full-batch Hamiltonian Monte Carlo achieve poor generalization under covariate shift, even underperforming classical estimation. We explain this surprising result, showing how a Bayesian model average can in fact be problematic under covariate shift, particularly in cases where linear dependencies in the input features cause a lack of posterior contraction. We additionally show why the same issue does not affect many approximate inference procedures, or classical maximum a-posteriori (MAP) training. Finally, we propose novel priors that improve the robustness of BNNs to many sources of covariate shift.
; ; ;
Award ID(s):
Publication Date:
Journal Name:
Advances in neural information processing systems
Sponsoring Org:
National Science Foundation
More Like this
  1. Reliable environmental context prediction is critical for wearable robots (e.g., prostheses and exoskeletons) to assist terrain-adaptive locomotion. This article proposed a novel vision-based context prediction framework for lower limb prostheses to simultaneously predict human's environmental context for multiple forecast windows. By leveraging the Bayesian neural networks (BNNs), our framework can quantify the uncertainty caused by different factors (e.g., observation noise, and insufficient or biased training) and produce a calibrated predicted probability for online decision-making. We compared two wearable camera locations (a pair of glasses and a lower limb device), independently and conjointly. We utilized the calibrated predicted probability for online decision-making and fusion. We demonstrated how to interpret deep neural networks with uncertainty measures and how to improve the algorithms based on the uncertainty analysis. The inference time of our framework on a portable embedded system was less than 80 ms/frame. The results in this study may lead to novel context recognition strategies in reliable decision-making, efficient sensor fusion, and improved intelligent system design in various applications.
  2. One of the central elements of any causal inference is an object called structural causal model (SCM), which represents a collection of mechanisms and exogenous sources of random variation of the system under investigation (Pearl, 2000). An important property of many kinds of neural networks is universal approximability: the ability to approximate any function to arbitrary precision. Given this property, one may be tempted to surmise that a collection of neural nets is capable of learning any SCM by training on data generated by that SCM. In this paper, we show this is not the case by disentangling the notions of expressivity and learnability. Specifically, we show that the causal hierarchy theorem (Thm. 1, Bareinboim et al., 2020), which describes the limits of what can be learned from data, still holds for neural models. For instance, an arbitrarily complex and expressive neural net is unable to predict the effects of interventions given observational data alone. Given this result, we introduce a special type of SCM called a neural causal model (NCM), and formalize a new type of inductive bias to encode structural constraints necessary for performing causal inferences. Building on this new class of models, we focus on solving twomore »canonical tasks found in the literature known as causal identification and estimation. Leveraging the neural toolbox, we develop an algorithm that is both sufficient and necessary to determine whether a causal effect can be learned from data (i.e., causal identifiability); it then estimates the effect whenever identifiability holds (causal estimation). Simulations corroborate the proposed approach.« less
  3. Abstract Deep neural networks (DNNs) have achieved state-of-the-art performance in many important domains, including medical diagnosis, security, and autonomous driving. In domains where safety is highly critical, an erroneous decision can result in serious consequences. While a perfect prediction accuracy is not always achievable, recent work on Bayesian deep networks shows that it is possible to know when DNNs are more likely to make mistakes. Knowing what DNNs do not know is desirable to increase the safety of deep learning technology in sensitive applications; Bayesian neural networks attempt to address this challenge. Traditional approaches are computationally intractable and do not scale well to large, complex neural network architectures. In this paper, we develop a theoretical framework to approximate Bayesian inference for DNNs by imposing a Bernoulli distribution on the model weights. This method called Monte Carlo DropConnect (MC-DropConnect) gives us a tool to represent the model uncertainty with little change in the overall model structure or computational cost. We extensively validate the proposed algorithm on multiple network architectures and datasets for classification and semantic segmentation tasks. We also propose new metrics to quantify uncertainty estimates. This enables an objective comparison between MC-DropConnect and prior approaches. Our empirical results demonstrate thatmore »the proposed framework yields significant improvement in both prediction accuracy and uncertainty estimation quality compared to the state of the art.« less
  4. Domain specific neural network accelerators have garnered attention because of their improved energy efficiency and inference performance compared to CPUs and GPUs. Such accelerators are thus well suited for resource-constrained embedded systems. However, mapping sophisticated neural network models on these accelerators still entails significant energy and memory consumption, along with high inference time overhead. Binarized neural networks (BNNs), which utilize single-bit weights, represent an efficient way to implement and deploy neural network models on accelerators. In this paper, we present a novel optical-domain BNN accelerator, named ROBIN , which intelligently integrates heterogeneous microring resonator optical devices with complementary capabilities to efficiently implement the key functionalities in BNNs. We perform detailed fabrication-process variation analyses at the optical device level, explore efficient corrective tuning for these devices, and integrate circuit-level optimization to counter thermal variations. As a result, our proposed ROBIN architecture possesses the desirable traits of being robust, energy-efficient, low latency, and high throughput, when executing BNN models. Our analysis shows that ROBIN can outperform the best-known optical BNN accelerators and many electronic accelerators. Specifically, our energy-efficient ROBIN design exhibits energy-per-bit values that are ∼4 × lower than electronic BNN accelerators and ∼933 × lower than a recently proposed photonic BNNmore »accelerator, while a performance-efficient ROBIN design shows ∼3 × and ∼25 × better performance than electronic and photonic BNN accelerators, respectively.« less
  5. De Vico Fallani, Fabrizio (Ed.)
    The exponential family random graph modeling (ERGM) framework provides a highly flexible approach for the statistical analysis of networks (i.e., graphs). As ERGMs with dyadic dependence involve normalizing factors that are extremely costly to compute, practical strategies for ERGMs inference generally employ a variety of approximations or other workarounds. Markov Chain Monte Carlo maximum likelihood (MCMC MLE) provides a powerful tool to approximate the maximum likelihood estimator (MLE) of ERGM parameters, and is generally feasible for typical models on single networks with as many as a few thousand nodes. MCMC-based algorithms for Bayesian analysis are more expensive, and high-quality answers are challenging to obtain on large graphs. For both strategies, extension to the pooled case—in which we observe multiple networks from a common generative process—adds further computational cost, with both time and memory scaling linearly in the number of graphs. This becomes prohibitive for large networks, or cases in which large numbers of graph observations are available. Here, we exploit some basic properties of the discrete exponential families to develop an approach for ERGM inference in the pooled case that (where applicable) allows an arbitrarily large number of graph observations to be fit at no additional computational cost beyond preprocessingmore »the data itself. Moreover, a variant of our approach can also be used to perform Bayesian inference under conjugate priors, again with no additional computational cost in the estimation phase. The latter can be employed either for single graph observations, or for observations from graph sets. As we show, the conjugate prior is easily specified, and is well-suited to applications such as regularization. Simulation studies show that the pooled method leads to estimates with good frequentist properties, and posterior estimates under the conjugate prior are well-behaved. We demonstrate the usefulness of our approach with applications to pooled analysis of brain functional connectivity networks and to replicated x-ray crystal structures of hen egg-white lysozyme.« less