Numerical tests on synthetic datasets evaluate the sparse GKDE, the stochastic generator, and the complete UQ-ANN framework, including out-of-sample predictions.
Abstract
Uncertainty quantification (UQ) is essential for reliable predictions in artificial neural networks when latent sources of randomness produce non-Gaussian outputs for a given deterministic input. We propose UQ-ANN, a data-driven framework composed of two coupled subnetworks. The first subnetwork contains two feedforward neural networks: ANN1 predicts the conditional mean and ANN2 the conditional covariance of a reduced representation obtained by principal component analysis (PCA) of the outputs. The second subnetwork is an analytical stochastic generator and requires no training. It is based on a sparse Gaussian kernel density estimator (GKDE) constructed by adaptive clustering. The sparsification replaces the kernels associated with the training points by a reduced set of cluster-based kernels, while third- and fourth-moment descriptors are included in the merging criterion to preserve non-Gaussian features. Independent realizations in the reduced space are then generated, transformed to reproduce the conditional mean and covariance predicted by the first subnetwork, and mapped back to the original output space. They provide conditional statistics, conditional probability density functions, and conditional prediction intervals. A marginal log-density diagnostic is also used to assess the stochastic approximation. Numerical tests on synthetic datasets evaluate the sparse GKDE, the stochastic generator, and the complete UQ-ANN framework, including out-of-sample predictions.
Optimal Transport Dropout (OTD) is introduced, which offers a way to learn structured predictive uncertainty without explicit posterior inference or ensembles of independently trained predictors.
Giacomo Lorenzon, Francesco Regazzoni· arXiv.org· 0 citations
Reliable uncertainty quantification (UQ) is essential for deploying neural networks in scientific and high-stakes applications, but full Bayesian inference over the network parameters is computationally infeasible. We propose a low-rank generalized Laplace approximation for neural-network UQ based on a small number of...
A capacity theorem shows that, under a full-rank feature assumption, a restricted subfamily matches the chaos coefficients of any Gaussian-latent random graph signal, with exponentially decaying truncation error under a growth condition; the task-level claims are established empirically.
Fred Xu, Thomas Markovich, Florence Regol et al.· arXiv.org· 0 citations
Stochastic neural network broadens deep learning from prediction‐oriented function approximation to a richer framework for statistical inference, and two related architectures are discussed: kernel‐expanded StoNet and sublinearly structured DNNs, which can be viewed as structural or theoretical extensions of the StoNet...
This paper introduces a novel algorithm for predicting conditional joint distributions of vector-valued targets in stochastic systems whose randomness is intrinsic rather than arising from observation errors or additive noise. Multivariate quantile regression also involves modeling conditional joint distributions but r...
A one-step-per-layer smoother that approximates as Gaussian only each layer's affine output is derived that applies both to deterministic systems with noisy observations and to stochastic systems described by output statistics.
Oren Wright, Hao-Ming Jing, Qiao-An Shen et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.