Dataset Information

Sampling-free Uncertainty Estimation in Gated Recurrent Units with Applications to Normative Modeling in Neuroimaging.

ABSTRACT: There has recently been a concerted effort to derive mechanisms in vision and machine learning systems to offer uncertainty estimates of the predictions they make. Clearly, there are benefits to a system that is not only accurate but also has a sense for when it is not. Existing proposals center around Bayesian interpretations of modern deep architectures - these are effective but can often be computationally demanding. We show how classical ideas in the literature on exponential families on probabilistic networks provide an excellent starting point to derive uncertainty estimates in Gated Recurrent Units (GRU). Our proposal directly quantifies uncertainty deterministically, without the need for costly sampling-based estimation. We show that while uncertainty is quite useful by itself in computer vision and machine learning, we also demonstrate that it can play a key role in enabling statistical analysis with deep networks in neuroimaging studies with normative modeling methods. To our knowledge, this is the first result describing sampling-free uncertainty estimation for powerful sequential models such as GRUs.

SUBMITTER: Hwang SJ

PROVIDER: S-EPMC7220220 | biostudies-literature | 2019 Jul

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Sampling-free Uncertainty Estimation in Gated Recurrent Units with Applications to Normative Modeling in Neuroimaging.

Hwang Seong Jae SJ Mehta Ronak R RR Kim Hyunwoo J HJ Johnson Sterling C SC Singh Vikas V

Uncertainty in artificial intelligence : proceedings of the ... conference. Conference on Uncertainty in Artificial Intelligence 20190701

There has recently been a concerted effort to derive mechanisms in vision and machine learning systems to offer uncertainty estimates of the predictions they make. Clearly, there are benefits to a system that is not only accurate but also has a sense for when it is not. Existing proposals center around Bayesian interpretations of modern deep architectures - these are effective but can often be computationally demanding. We show how classical ideas in the literature on exponential families on pro ...[more]

PMID: 32405271

Similar Datasets

Project description:Intraclass correlation (ICC) is a reliability metric that gauges similarity when, for example, entities are measured under similar, or even the same, well-controlled conditions, which in MRI applications include runs/sessions, twins, parent/child, scanners, sites, and so on. The popular definitions and interpretations of ICC are usually framed statistically under the conventional ANOVA platform. Here, we provide a comprehensive overview of ICC analysis in its prior usage in neuroimaging, and we show that the standard ANOVA framework is often limited, rigid, and inflexible in modeling capabilities. These intrinsic limitations motivate several improvements. Specifically, we start with the conventional ICC model under the ANOVA platform, and extend it along two dimensions: first, fixing the failure in ICC estimation when negative values occur under degenerative circumstance, and second, incorporating precision information of effect estimates into the ICC model. These endeavors lead to four modeling strategies: linear mixed-effects (LME), regularized mixed-effects (RME), multilevel mixed-effects (MME), and regularized multilevel mixed-effects (RMME). Compared to ANOVA, each of these four models directly provides estimates for fixed effects and their statistical significances, in addition to the ICC estimate. These new modeling approaches can also accommodate missing data and fixed effects for confounding variables. More importantly, we show that the MME and RMME approaches offer more accurate characterization and decomposition among the variance components, leading to more robust ICC computation. Based on these theoretical considerations and model performance comparisons with a real experimental dataset, we offer the following general-purpose recommendations. First, ICC estimation through MME or RMME is preferable when precision information (i.e., weights that more accurately allocate the variances in the data) is available for the effect estimate; when precision information is unavailable, ICC estimation through LME or the RME is the preferred option. Second, even though the absolute agreement version, ICC(2,1), is presently more popular in the field, the consistency version, ICC(3,1), is a practical and informative choice for whole-brain ICC analysis that achieves a well-balanced compromise when all potential fixed effects are accounted for. Third, approaches for clear, meaningful, and useful result reporting in ICC analysis are discussed. All models, ICC formulations, and related statistical testing methods have been implemented in an open source program 3dICC, which is publicly available as part of the AFNI suite. Even though our work here focuses on the whole-brain level, the modeling strategy and recommendations can be equivalently applied to other situations such as voxel, region, and network levels.

Project description:Taking parameter uncertainty into account is key to make drug development decisions such as testing whether trial endpoints meet defined criteria. Currently used methods for assessing parameter uncertainty in NLMEM have limitations, and there is a lack of diagnostics for when these limitations occur. In this work, a method based on sampling importance resampling (SIR) is proposed, which has the advantage of being free of distributional assumptions and does not require repeated parameter estimation. To perform SIR, a high number of parameter vectors are simulated from a given proposal uncertainty distribution. Their likelihood given the true uncertainty is then approximated by the ratio between the likelihood of the data given each vector and the likelihood of each vector given the proposal distribution, called the importance ratio. Non-parametric uncertainty distributions are obtained by resampling parameter vectors according to probabilities proportional to their importance ratios. Two simulation examples and three real data examples were used to define how SIR should be performed with NLMEM and to investigate the performance of the method. The simulation examples showed that SIR was able to recover the true parameter uncertainty. The real data examples showed that parameter 95 % confidence intervals (CI) obtained with SIR, the covariance matrix, bootstrap and log-likelihood profiling were generally in agreement when 95 % CI were symmetric. For parameters showing asymmetric 95 % CI, SIR 95 % CI provided a close agreement with log-likelihood profiling but often differed from bootstrap 95 % CI which had been shown to be suboptimal for the chosen examples. This work also provides guidance towards the SIR workflow, i.e.,which proposal distribution to choose and how many parameter vectors to sample when performing SIR, using diagnostics developed for this purpose. SIR is a promising approach for assessing parameter uncertainty as it is applicable in many situations where other methods for assessing parameter uncertainty fail, such as in the presence of small datasets, highly nonlinear models or meta-analysis.

Dataset Information

Sampling-free Uncertainty Estimation in Gated Recurrent Units with Applications to Normative Modeling in Neuroimaging.

Publications

Sampling-free Uncertainty Estimation in Gated Recurrent Units with Applications to Normative Modeling in Neuroimaging.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets