Changes In Model Output Research Articles

Improving Adversarial Robustness to Sensitivity and Invariance Attacks with Deep Metric Learning (Student Abstract)

Intentionally crafted adversarial samples have effectively exploited weaknesses in deep neural networks. A standard method in adversarial robustness assumes a framework to defend against samples crafted by minimally perturbing a sample such that its corresponding model output changes. These sensitivity attacks exploit the model's sensitivity toward task-irrelevant features. Another form of adversarial sample can be crafted via invariance attacks, which exploit the model underestimating the importance of relevant features. Previous literature has indicated a tradeoff in defending against both attack types within a strictly L-p bounded defense. To promote robustness toward both types of attacks beyond Euclidean distance metrics, we use metric learning to frame adversarial regularization as an optimal transport problem. Our preliminary results indicate that regularizing over invariant perturbations in our framework improves both invariant and sensitivity defense.

Explaining Black Box Drug Target Prediction Through Model Agnostic Counterfactual Samples.

Many high-performance DTA deep learning models have been proposed, but they are mostly black-box and thus lack human interpretability. Explainable AI (XAI) can make DTA models more trustworthy, and allows to distill biological knowledge from the models. Counterfactual explanation is one popular approach to explaining the behaviour of a deep neural network, which works by systematically answering the question "How would the model output change if the inputs were changed in this way?". We propose a multi-agent reinforcement learning framework, Multi-Agent Counterfactual Drug-target binding Affinity (MACDA), to generate counterfactual explanations for the drug-protein complex. Our proposed framework provides human-interpretable counterfactual instances while optimizing both the input drug and target for counterfactual generation at the same time. We benchmark the proposed MACDA framework using the Davis and PDBBind dataset and find that our framework produces more parsimonious explanations with no loss in explanation validity, as measured by encoding similarity. We then present a case study involving ABL1 and Nilotinib to demonstrate how MACDA can explain the behaviour of a DTA model in the underlying substructure interaction between inputs in its prediction, revealing mechanisms that align with prior domain knowledge.

Climate and parameter sensitivity and induced uncertainties in carbon stock projections for European forests (using LPJ-GUESS 4.0)

Abstract. Understanding uncertainties and sensitivities of projected ecosystem dynamics under environmental change is of immense value for research and climate change policy. Here, we analyze sensitivities (change in model outputs per unit change in inputs) and uncertainties (changes in model outputs scaled to uncertainty in inputs) of vegetation dynamics under climate change, projected by a state-of-the-art dynamic vegetation model (LPJ-GUESS v4.0) across European forests (the species Picea abies, Fagus sylvatica and Pinus sylvestris), considering uncertainties of both model parameters and environmental drivers. We find that projected forest carbon fluxes are most sensitive to photosynthesis-, water-, and mortality-related parameters, while predictive uncertainties are dominantly induced by environmental drivers and parameters related to water and mortality. The importance of environmental drivers for predictive uncertainty increases with increasing temperature. Moreover, most of the interactions of model inputs (environmental drivers and parameters) are between environmental drivers themselves or between parameters and environmental drivers. In conclusion, our study highlights the importance of environmental drivers not only as contributors to predictive uncertainty in their own right but also as modifiers of sensitivities and thus uncertainties in other ecosystem processes. Reducing uncertainty in mortality-related processes and accounting for environmental influence on processes should therefore be a focus in further model development.

Influence of modifications (from AoB2015 to v0.5) in the Vegetation Optimality Model

Abstract. The Vegetation Optimality Model (VOM, Schymanski et al., 2009, 2015) is an optimality-based, coupled water–vegetation model that predicts vegetation properties and behaviour based on optimality theory rather than calibrating vegetation properties or prescribing them based on observations, as most conventional models do. Several updates to previous applications of the VOM have been made for the study in the accompanying paper of Nijzink et al. (2022), where we assess whether optimality theory can alleviate common shortcomings of conventional models, as identified in a previous model inter-comparison study along the North Australian Tropical Transect (NATT, Whitley et al., 2016). Therefore, we assess in this technical paper how the updates to the model and input data would have affected the original results of Schymanski et al. (2015), and we implemented these changes one at a time. The model updates included extended input data, the use of variable atmospheric CO2 levels, modified soil properties, implementation of free drainage conditions, and the addition of grass rooting depths to the optimized vegetation properties. A systematic assessment of these changes was carried out by adding each individual modification to the original version of the VOM at the flux tower site of Howard Springs, Australia. The analysis revealed that the implemented changes affected the simulation of mean annual evapotranspiration (ET) and gross primary productivity (GPP) by no more than 20 %, with the largest effects caused by the newly imposed free drainage conditions and modified soil texture. Free drainage conditions led to an underestimation of ET and GPP in comparison with the results of Schymanski et al. (2015), whereas more fine-grained soil textures increased the water storage in the soil and resulted in increased GPP. Although part of the effect of free drainage was compensated for by the updated soil texture, when combining all changes, the resulting effect on the simulated fluxes was still dominated by the effect of implementing free drainage conditions. Eventually, the relative error for the mean annual ET, in comparison with flux tower observations, changed from an 8.4 % overestimation to an 10.2 % underestimation, whereas the relative errors for the mean annual GPP remained similar, with an overestimation that slightly reduced from 17.8 % to 14.7 %. The sensitivity to free drainage conditions suggests that a realistic representation of groundwater dynamics is very important for predicting ET and GPP at a tropical open-forest savanna site as investigated here. The modest changes in model outputs highlighted the robustness of the optimization approach that is central to the VOM architecture.

Elucidating the Common Basis for Task-Dependent Differential Manifestations of Category Advantage: A Decision Theoretic Approach.

Cross-category hues are differentiated easier than otherwise equidistant hues that belong to the same linguistic category. This effect is typically manifested through both accuracy and response time gains in tasks with a memory component, whereas only response times are affected when there is no memory component. This raises the question of whether there is a common generative process underlying the differential behavioral manifestations of category advantage in color perception. For instance, within the framework of noisy evidence accumulation models, changes in accuracy can be readily attributed to an increase in the efficacy of perceptual evidence integration (after controlling for threshold setting), whereas changes in response time can also be attributed to shorter nondecisional delays (e.g., due to facilitated signal detection). To address the latent decision processes underlying category advantage across different behavioral demands, we introduce a decision-theoretic perspective (i.e., diffusion decision model) to categorical color perception in three complementary experiments. In Experiment 1, we collected data from a binary color naming task (1) to determine the green-blue boundary in our sample and (2) to trace how parameter estimates of interest in the model output change as a function of color typicality. In Experiments 2 and 3, we used same-different task paradigms (with and without a memory component, respectively) and traced the category advantage in color discrimination in two parameters of the diffusion decision model: nondecision time and drift rate. An increase in drift rate predominantly characterized the category advantage in both tasks. Our results show that improved efficiency in perceptual evidence integration is a common driving force behind different manifestations of category advantage.

Stock assessment on fishery‐dependent data: Effect of data quality and parametrisation for a red snapper fishery

AbstractData availability, and unreported and unregulated fishing are significant obstacles to evaluating stock status, especially in tropical areas. Limitations in data quantity and quality can lead to model misspecification and erroneous data treatments, potentially causing important changes in model outputs and subsequent management implications. Red snapper Lutjanus purpureus (Poey) in French Guiana provides an example of a stock with a long‐time series of fishery‐dependent data subject to large uncertainty. A flexible catch‐at‐age model (Stock Synthesis) was applied to the available data and compared to an historically applied assessment approach. Inter‐model variability based on different model specifications and data treatments were compared to identify better the status of the resource. Results showed that a major source of uncertainty in the model was the inclusion of a catch‐per‐unit‐effort abundance index with questionable ability to track abundance. The Stock Synthesis model provided a more flexible and viable method than the virtual population analysis approach. Despite large uncertainty, models depicted a similar trend with a notable stock depletion in the late 1990s but with two distinct biomass trends in more recent years depending on the treatment. To reduce uncertainty and preserve this important economic resource, new data collection programmes and management policies are needed.

A missing source of uncertainty: forcing-dependent model parameter sensitivity

Climate modelling groups usually work with a handful of model versions (parameter combinations) that reproduce certain targeted aspects of observed climate within a certain validity range and apply them for studying future climate change. What is of concern is whether these retained model versions, with respect to their de-selected counterparts, continue being optimal for future climate that is supposed to distinctly differ from the present one. Extrapolating model performances beyond their validity range requires model parameter sensitivity (i.e., changes in model output due to changes in model parameters) remains more or less stationary despite different forcing conditions. This requirement, however, is shown to be ill-grounded by exemplified analyses of resolution sensitivity in an Earth System Model under different forcing conditions, whereby model resolution is handled as a model parameter in a wider sense: (i) Model resolution sensitivity depends on the forcing conditions applied; moreover, (ii) The further the forcing deviates from a reference state, the earlier one can detect a systematic change in model resolution sensitivity, in particular, in its spatial details. Therefore, model parameter sensitivity and forcing conditions should be evaluated as a compound; failure to account for this relation leads to a systematic underestimation of uncertainty in forced responses of climate models, thereby imposing hazardous impacts on practical applications of CMIP outputs.

Active and Incremental Learning with Weak Supervision

Large amounts of labeled training data are one of the main contributors to the great success that deep models have achieved in the past. Label acquisition for tasks other than benchmarks can pose a challenge due to requirements of both funding and expertise. By selecting unlabeled examples that are promising in terms of model improvement and only asking for respective labels, active learning can increase the efficiency of the labeling process in terms of time and cost. In this work, we describe combinations of an incremental learning scheme and methods of active learning. These allow for continuous exploration of newly observed unlabeled data. We describe selection criteria based on model uncertainty as well as expected model output change (EMOC). An object detection task is evaluated in a continuous exploration context on the PASCAL VOC dataset. We also validate a weakly supervised system based on active and incremental learning in a real-world biodiversity application where images from camera traps are analyzed. Labeling only 32 images by accepting or rejecting proposals generated by our method yields an increase in accuracy from 25.4 to 42.6%.

A polynomial chaos framework for probabilistic predictions of storm surge events

We present a polynomial chaos-based framework to quantify the uncertainties in predicting hurricane-induced storm surges. Perturbation strategies are proposed to characterize poorly known time-dependent input parameters, such as tropical cyclone track and wind as well as space-dependent bottom stresses, using a handful of stochastic variables. The input uncertainties are then propagated through an ensemble calculation and a model surrogate is constructed to represent the changes in model output caused by changes in the model input. The statistical analysis is then performed using the model surrogate once its reliability has been established. The procedure is illustrated by simulating the flooding caused by Hurricane Gustav 2008 using the ADvanced CIRCulation model. The hurricane’s track and intensity are perturbed along with the bottom friction coefficients. A sensitivity analysis suggests that the track of the tropical cyclone is the dominant contributor to the peak water level forecast, while uncertainties in wind speed and in the bottom friction coefficient show minor contributions. Exceedance probability maps with different levels are also estimated to identify the most vulnerable areas.

Parametric sensitivity analysis as an essential ingredient of spectroscopic ellipsometry data modeling: An application of the Morris screening method

Parametric sensitivity analysis (SA) is an essential tool in optical data analysis aiming to quantify the relative importance of optical model parameters and identify those with a low influence which should remain fixed during analysis in order to simplify a model. Typically, such sensitivity studies are performed using a local sensitivity analysis, also referred to as “one-at-a-time” analysis, which estimates parameter sensitivity by evaluating the changes in model output while perturbing one parameter at a time and holding all other parameters fixed. All the pitfalls of the local sensitivity approach are quite obvious and well known; however, local SA is still prevailing in the area of data analysis for spectroscopic ellipsometry. In this study, we discuss applications of a global qualitative analysis introduced by Morris in 1991 which determines an overall significance for each model parameter and describes its nonlinear effect on model’s output and/or interactions with other parameters. The Morris method allows effective separation of the model parameters into different groups such as parameters with negligible effects, parameters with large linear effects without interactions, and parameters with large nonlinear and/or interaction effects. Here, we carried out the Morris sensitivity analysis to reveal the model parameter significance ranking in a test case of the B-spline model for titanium nitride (TiN). Such a screening approach can be considered as a practical tool for the initial global SA strategy prior to true “all-at-a-time” global sensitivity analysis.

An Active Learning Framework for Alpha Matting

Good trimap is essential for high-quality alpha matte. However, making high-quality trimap is hardwork, especially for complex images. In this paper, an active learning framework is proposed to make high quality trimap. There are two active learning methods which are employed: minimization of uncertainty sampling (MUS) and maximization of expected model output change (EMOC). MUS model finds the informative area in image which can decrease the uncertain sampling of alpha matte. EMOC model finds the important areas in image which can give the maximum expected output change of alpha matte. Two methods are combined to define the active map. Active map shows important areas which are informative in image. It can help users to make high quality trimap. The analysis and evaluation of benchmark datasets show that proposed method is effective.

Updating a gas dynamics model using estimates for California sea lions (Zalophus californianus)

Theoretical models are used to predict how breath-hold diving vertebrates manage O2, CO2, and N2 while underwater. One recent gas dynamics model used available lung and tracheal compliance data from various species. As variation in respiratory compliance significantly affects alveolar compression and pulmonary shunt, the current study objective was to evaluate changes in model output when using species-specific parameters from California sea lions (Zalophus californianus). We explored the effects of lung and dead space compliance on the uptake of N2, O2, and CO2 in various tissues during a series of hypothetical dives. The updated parameters allowed for increased compliance of the lungs and an increased stiffness in the trachea. When comparing updated model output with a model using previous compliance values, there was a large decrease in N2 uptake but little change in O2 and CO2 levels. Therefore, previous models may overestimate N2 tensions and the risk of gas-related disease, such as decompression sickness (DCS), in marine mammals.

Forecasting Sovereign Default Risk with Merton’s Model

Merton’s structural model for sovereigns is proven to be useful to analyze the default risk of a country. We are the first to investigate how fast CDS spreads react to changes in model inputs and outputs. CDS spread changes strongly correlate with exchange rate returns, which are an input to the model. But CDS spread changes on average react with a delay to changes in model outputs such as the distance to default, the default probability, and model spreads. Hence, contingency claim analysis for sovereigns provides useful predictions for CDS spreads. <b>TOPICS:</b>Fixed income and structured finance, credit default swaps

Computational Sensitivity Analysis on a Mathematical Model of Epileptic Seizures

Temporal lobe epilepsy is a serious neurological disorder characterized by complex partial seizures, which are thought to originate in the hippocampus.В Ordinary differential equation modeling has been used to describe changes in membrane potential of excitatory and inhibitory cells in order to gain insight into seizure propagation.В In the current study, a system of ordinary differential equations based on previous modeling is used with distinct biologically reasonable values for membrane capacitance in order to determine model sensitivity to that parameter. Because delay differential equations are used in the model, sensitivity is investigated computationally by examining the variation in output relative to various inputs. Membrane capacitance was found to affect model predictions and whether groups of cells exhibited the same behavior after a certain period of time.В Hence, membrane capacitance is a critical parameter when modeling changes in membrane potential and should be incorporated clearly.В Changes in model output as a result of changes in a time delay parameter are also investigated.

Calculating NWoW office space with the PACT model

Purpose – The purpose of this paper is to illustrate the working of the PACT calculation model, a tool to determine office space dimensions. New ways of working (NWoW) seem to have become a fixed value in facility management (FM) practice in The Netherlands today. Stimulated by new technological possibilities, companies are rethinking their office environments to make workplaces more flexible and their use “activity related”. However, this requires a different approach to quantify the needed space and determine the types of workplaces to fit organizations’ processes. The PLaces and ACTivities (PACT) calculation model allows (facility) managers to gain an insight in the number and type of spaces needed, modulated by different scenarios and fitting to the organization and its work processes. Design/methodology/approach – This article mainly aims to present the PACT model: an office space calculation tool. A case study is presented and calculated to compare an actual work environment of an organization to the PACT calculated results. As input for the model, data were used that were available before the work environment changes in 2007. Additionally, one scenario of a different workplace use is calculated which helped to visualize the accuracy and validity of the model. Findings – When comparing the post hoc PACT calculated space to the real-life work environment, the number of calculated workplaces and the ratio to the number of employees do not seem to differ strongly. However, substantially less meeting space is calculated by the model, and some elements might require more testing to verify it completely. The scenario calculation shows that the model output changes to adapt to a more flexible work process. Originality/value – Even though calculation and simulation models for office space are available, the described model puts together many different elements to provide a more holistic calculation. Elements like, for instance, absence, activities and occupational choices are combined.

Investigation of Uncertainty Changes in Model Outputs for Finite-Element Model Updating Using Structural Health Monitoring Data

This article aims to investigate the effect of uncertainties on the predicted response of structures using updated finite-element models (FEMs). Modeling uncertainties are quantified by fuzzy numbers and are incorporated into the fuzzy FEM updating procedure. The impact of the amount and types of data used on the performance of the updated model is investigated. In order to perform the complex FEM updating calculations, which generally take too much time for complex models, a Gaussian process (GP) is used as a surrogate model. The central composite design (CCD) method is used to sample the input parameter space for more accurate GP models. Genetic algorithms (GA) are employed to solve the inverse fuzzy model updating problem. Additional constraints are presented to capture the variation space of the uncertain response parameters. The University of Central Florida benchmark test structure, which is designed to represent short-span to medium-span bridges, is used in the scope of uncertainty quantification study. Static and dynamic experimental test data obtained from the benchmark structure under different loadings and conditions are used for the demonstration. A damage case, in which the stiffness reduction in boundaries is simulated by using flexible pads, is considered. The results show that appropriate data sets, which contain the least uncertainty, should be generated instead of involving the entire set of measurements obtained from different tests. Nevertheless, uncertainty quantification should be employed to find the variation range of uncertain responses predicted by simplified FEM models.

A Study of Interactions in the Risk Assessment of Complex Engineering Systems: An Application to Space PSA

Risk managers are often confronted with the evaluation of operational policies in which two or more system components are simultaneously affected by a change. In these instances, the decision-making process should be informed by the relevance of interactions. However, because of system and model complexity, a rigorous study for determining whether and how interactions quantitatively impact operational choices has not been developed yet. In light of the central role played by the multilinearity of the decision support models, we investigate the presence of interactions in multilinear functions first. We identify interactions that can be a priori excluded from the analysis. We introduce sensitivity measures that apportion the model output change to individual factors and interaction contributions in an exact fashion. The sensitivity measures are linked to graphical representation methods as tornado diagrams and Pareto charts, and a systematic way of inferring managerial insights is presented. We then specialize the findings to reliability and probabilistic safety assessment (PSA) problems. We set forth a procedure for determining the magnitude of changes that make interactions relevant in the analysis. Quantitative results are discussed by application to a PSA model developed at NASA to support decision making in space mission planning and design. Numerical findings show that suboptimal decisions concerning the components on which to focus managerial attention can be made, if the decision-making process is not informed by the relevance of interactions.

Explaining and Predicting Sovereign Credit Risk with Exchange Rate Volatility

Merton's structural model for sovereigns is proven to be useful to analyze the default risk of a country. We are the first to investigate how fast CDS spreads react to changes in model inputs and outputs. CDS spread changes strongly correlate with exchange rate returns, which are an input to the model. But CDS spread changes on average react with a delay to changes in model outputs such as the distance to default, the default probability and model spreads. Hence contingency claim analysis for sovereigns provides useful predictions for CDS spreads.

Associating changes in output behavior with changes in parameter values in spiking and bursting neuron models

Several recent studies have demonstrated that neuronal models allow multiple parameter value solutions for a given output. In the face of this variability of parameter values, what can be learned about neural function through parameter value differences? Here, in two different models, we examine this question by attempting to reconstruct the source of model output changes based on simple statistical analyses of parameter distributions generated by automated searches. We conclude that changes to parameter values or their associated distributions do not reliably reflect the specific mechanisms responsible for a given change in output.

Adjoint Sensitivity Analysis of Contaminant Concentrations in Water Distribution Systems

Sensitivity analysis is used to determine how a system state or a model output changes due to a change in the value of a system parameter or a model input. We present the adjoint approach for determining the sensitivity of the concentration of a contaminant in a water distribution system to a change in a system parameter such as the location of the source of contamination, the reaction rate of the contaminant, and others. With the adjoint method, the sensitivity of the model output to any number of parameters can be obtained with one simulation of the adjoint model. If the number of parameters of interest exceeds the number of model outputs for which the sensitivity is desired, the adjoint method is more efficient than traditional direct methods of calculating sensitivities. We develop the adjoint equations for water quality in a water distribution system, verify the adjoint-based sensitivity equation using an analytical example, and demonstrate the numerical calculation of adjoint sensitivities using EPANET.