Bayesian Analysis of State Voter Registration Database Integrity

How do we ensure a statewide voter registration database’s accuracy and integrity, especially when the database depends on aggregating decentralized, sub-state data with different list maintenance practices? We develop a Bayesian multivariate multilevel model to account for correlated patterns of change over time in multiple response variables, and label statewide anomalies using deviations from model predictions. We apply our model to California’s 22 million registered voters, using 25 snapshots from the 2020 presidential election. We estimate countywide change rates for multiple response variables such as changes in voter’s partisan affiliation and jointly model these changes. The model outperforms a simple interquartile range (IQR) detection when tested with synthetic data. This is a proof-of-concept that demonstrates the utility of the Bayesian methodology, as despite the heterogeneity in list maintenance practices, a principled, statistical approach is useful. At the county level, the total numbers of anomalies are positively correlated with the average election cost per registered voter between 2017 and 2019. Given the recent efforts to modernize and secure voter list maintenance procedures in the For the People Act of 2021, we argue that checking whether counties or municipalities are behaving similarly at the state level is also an essential step in ensuring electoral integrity.

Assessing the value meteorological ensembles add to dispersion modelling using hypothetical releases

Atmospheric dispersion model output is frequently used to provide advice to decision makers, for example, about the likely location of volcanic ash erupted from a volcano or the location of deposits of radioactive material released during a nuclear accident. Increasingly, scientists and decision makers are requesting information on the uncertainty of these dispersion model predictions. One source of uncertainty is in the meteorology used to drive the dispersion model, and in this study ensemble meteorology from the Met Office ensemble prediction system is used to provide meteorological uncertainty to dispersion model predictions. Two hypothetical scenarios, one volcanological and one radiological, are repeated every 12 h over a period of 4 months. The scenarios are simulated using ensemble meteorology and deterministic forecast meteorology and compared to output from simulations using analysis meteorology using the Brier skill score. Adopting the practice commonly used in evaluating numerical weather prediction (NWP) models where observations are sparse or non-existent, we consider output from simulations using analysis NWP data to be truth. The results show that on average the ensemble simulations perform better than the deterministic simulations, although not all individual ensemble simulations outperform their deterministic counterpart. The results also show that greater skill scores are achieved by the ensemble simulation for later time steps rather than earlier time steps. In addition there is a greater increase in skill score over time for deposition than for air concentration. For the volcanic ash scenarios it is shown that the performance of the ensemble at one flight level can be different to that at a different flight level; e.g. a negative skill score might be obtained for FL350-550 and a positive skill score for FL200-350. This study does not take into account any source term uncertainty, but it does take the first steps towards demonstrating the value of ensemble dispersion model predictions.

Exogenous capture accounts for fundamental differences between prosaccade and antisaccade performance

Oculomotor circuits generate eye movements based on the physical salience of objects and current behavioral goals, exogenous and endogenous influences, respectively. However, the interactions between exogenous and endogenous mechanisms and their dynamic contributions to target selection have been difficult to resolve because they evolve extremely rapidly. In a recent study (Salinas et al., 2019), we achieved the necessary temporal precision using an urgent variant of the antisaccade task wherein motor plans are initiated early and choice accuracy depends sharply on when exactly the visual cue information becomes available. Empirical and modeling results indicated that the exogenous signal arrives ~80 ms after cue onset and rapidly accelerates the (incorrect) plan toward the cue, whereas the informed endogenous signal arrives ~25 ms later to favor the (correct) plan away from the cue. Here, we scrutinize a key mechanistic hypothesis about this dynamic, that the exogenous and endogenous signals act at different times and independently of each other. We test quantitative model predictions by comparing the performance of human participants instructed to look toward a visual cue versus away from it under high urgency. We find that, indeed, the exogenous response is largely impervious to task instructions; it simply flips its sign relative to the correct choice, and this largely explains the drastic differences in psychometric performance between the two tasks. Thus, saccadic choices are strongly dictated by the alignment between salience and behavioral goals.

Economical defence of resources structures territorial space use in a cooperative carnivore

Ecologists have long sought to understand space use and mechanisms underlying patterns observed in nature. We developed an optimality landscape and mechanistic territory model to understand mechanisms driving space use and compared model predictions to empirical reality. We demonstrate our approach using grey wolves ( Canis lupus ). In the model, simulated animals selected territories to economically acquire resources by selecting patches with greatest value, accounting for benefits, costs and trade-offs of defending and using space on the optimality landscape. Our approach successfully predicted and explained first- and second-order space use of wolves, including the population's distribution, territories of individual packs, and influences of prey density, competitor density, human-caused mortality risk and seasonality. It accomplished this using simple behavioural rules and limited data to inform the optimality landscape. Results contribute evidence that economical territory selection is a mechanistic bridge between space use and animal distribution on the landscape. This approach and resulting gains in knowledge enable predicting effects of a wide range of environmental conditions, contributing to both basic ecological understanding of natural systems and conservation. We expect this approach will demonstrate applicability across diverse habitats and species, and that its foundation can help continue to advance understanding of spatial behaviour.

Training calibration-based counterfactual explainers for deep learning models in medical image analysis

The rapid adoption of artificial intelligence methods in healthcare is coupled with the critical need for techniques to rigorously introspect models and thereby ensure that they behave reliably. This has led to the design of explainable AI techniques that uncover the relationships between discernible data signatures and model predictions. In this context, counterfactual explanations that synthesize small, interpretable changes to a given query while producing desired changes in model predictions have become popular. This under-constrained, inverse problem is vulnerable to introducing irrelevant feature manipulations, particularly when the model’s predictions are not well-calibrated. Hence, in this paper, we propose the TraCE (training calibration-based explainers) technique, which utilizes a novel uncertainty-based interval calibration strategy for reliably synthesizing counterfactuals. Given the wide-spread adoption of machine-learned solutions in radiology, our study focuses on deep models used for identifying anomalies in chest X-ray images. Using rigorous empirical studies, we demonstrate the superiority of TraCE explanations over several state-of-the-art baseline approaches, in terms of several widely adopted evaluation metrics. Our findings show that TraCE can be used to obtain a holistic understanding of deep models by enabling progressive exploration of decision boundaries, to detect shortcuts, and to infer relationships between patient attributes and disease severity.

Humans optimally anticipate and compensate for an uneven step during walking

The simple task of walking up a sidewalk curb is actually a dynamic prediction task. The curb is a disturbance that could cause a loss of momentum if not anticipated and compensated for. It might be possible to adjust momentum sufficiently to ensure undisturbed time of arrival, but there are infinite possible ways to do so. Much of steady, level gait is determined by energy economy, which should be at least as important with terrain disturbances. It is, however, unknown whether economy also governs walking up a curb, and whether anticipation helps. Here we show that humans compensate with an anticipatory pattern of forward speed adjustments, predicted by a criterion of minimizing mechanical energy input. The strategy is mechanistically predicted by optimal control for a simple model of bipedal walking dynamics, with each leg's push-off work as input. Optimization predicts a tri-phasic trajectory of speed (and thus momentum) adjustments, including an anticipatory phase. In experiment, human subjects ascend an artificial curb with the predicted tri-phasic trajectory, which approximately conserves overall walking speed relative to undisturbed flat ground. The trajectory involves speeding up in a few steps before the curb, losing considerable momentum from ascending it, and then regaining speed in a few steps thereafter. Descending the curb entails a nearly opposite, but still anticipatory, speed fluctuation trajectory, in agreement with model predictions that speed fluctuation amplitudes should scale linearly with curb height. The fluctuation amplitudes also decrease slightly with faster average speeds, also as predicted by model. Humans can reason about the dynamics of walking to plan anticipatory and economical control, even with a sidewalk curb in the way.

Finite deformation and fractional order viscoelasticity of an auxetic foam

Auxetic foams exhibit novel mechanical properties due to their unique microstructure for improved energy-absorption and cavity expansion applications that have fascinated the scientific community since their inception. Given the advancements in material processing and performance of polymer open cell auxetic foams, there is a strong desire to fully understand the nonlinear rate-dependent deformation of these materials. The influence of nonlinear compressibility is introduced here along with relaxation effects to improve model predictions for different stretch rates and finite deformation regimes. The viscoelastic behavior of the material is analyzed by comparing fractional order and integer order calculus models. All results are statistically validated using maximum entropy methods to obtain Bayesian posterior densities for the hyperelastic, auxetic, and viscoelastic parameters. It is shown that fractional order viscoelasticity provides [Formula: see text]–[Formula: see text] improvement in prediction over integer order viscoelastic models when the model is calibrated at higher stretch rates where viscoelasticity is more significant.

Achieving 90% In Data-Centric Industry Deep Learning Task

In industry deep learning application, our manually labeled data has a certain number of noisy data. To solve this problem and achieve more than 90 score in dev dataset, we present a simple method to find the noisy data and re-label the noisy data by human, given the model predictions as references in human labeling. In this paper, we illustrate our idea for a broad set of deep learning tasks, includes classification, sequence tagging, object detection, sequence generation, click-through rate prediction. The experimental results and human evaluation results verify our idea.

Monitoring and Modeling of Saline-Sodic Vertisol Reclamation by Echinochloa stagnina

Soil salinity due to irrigation is a major constraint to agriculture, particularly in arid and semi-arid zones, due to water scarcity and high evaporation rates. Reducing salinity is a fundamental objective for protecting the soil and supporting agricultural production. The present study aimed to empirically measure and simulate with a model, the reduction in soil salinity in a Vertisol by the cultivation and irrigation of Echinochloa stagnina. Laboratory soil column experiments were conducted to test three treatments: (i) ponded bare soil without crops, (ii) ponded soil cultivated with E. stagnina in two successive cropping seasons and (iii) ponded soil permanently cultivated with E. stagnina with a staggered harvest. After 11 months of E. stagnina growth, the electrical conductivity of soil saturated paste (ECe) decreased by 79–88% in the topsoil layer (0–8 cm) in both soils cultivated with E. stagnina and in bare soil. In contrast, in the deepest soil layer (18–25 cm), the ECe decreased more in soil cultivated with E. stagnina (41–83%) than in bare soil (32–58%). Salt stocks, which were initially similar in the columns, decreased more in soil cultivated with E. stagnina (65–87%) than in bare soil (34–45%). The simulation model Hydrus-1D was used to predict the general trends in soil salinity and compare them to measurements. Both the measurements and model predictions highlighted the contrast between the two cropping seasons: soil salinity decreased slowly during the first cropping season and rapidly during the second cropping season following the intercropping season. Our results also suggested that planting E. stagnina was a promising option for controlling the salinity of saline-sodic Vertisols.