Undesirable Bias Research Articles

FedART: A neural model integrating federated learning and adaptive resonance theory

Federated Learning (FL) has emerged as a promising paradigm for collaborative model training across distributed clients while preserving data privacy. However, prevailing FL approaches aggregate the clients’ local models into a global model through multi-round iterative parameter averaging. This leads to the undesirable bias of the aggregated model towards certain clients in the presence of heterogeneous data distributions among the clients. Moreover, such approaches are restricted to supervised classification tasks and do not support unsupervised clustering. To address these limitations, we propose a novel one-shot FL approach called Federated Adaptive Resonance Theory (FedART) which leverages self-organizing Adaptive Resonance Theory (ART) models to learn category codes, where each code represents a cluster of similar data samples. In FedART, the clients learn to associate their private data with various local category codes. Under heterogeneity, the local codes across different clients represent heterogeneous data. In turn, a global model takes these local codes as inputs and aggregates them into global category codes, wherein heterogeneous client data is indirectly represented by distinctly encoded global codes, in contrast to the averaging out of parameters in the existing approaches. This enables the learned global model to handle heterogeneous data. In addition, FedART employs a universal learning mechanism to support both federated classification and clustering tasks. Our experiments conducted on various federated classification and clustering tasks show that FedART consistently outperforms state-of-the-art FL methods on data with heterogeneous distribution across clients.

Bayesian evidence for the neural dissociation between finger and hand imitation skills

Abstract For limb apraxia ‒ a heterogeneous disorder of higher motor cognition following stroke ‒ an enduring debate has arisen regarding the existence of dissociating neural correlates for finger and hand gestures in the left hemisphere. We re-assessed this question asking whether previous attempts analysing pooled samples of patients with deficits in only one and patients with deficits in both imitation types might have led to systematically biased results. We conducted frequentist and Bayesian, voxelwise, and regionwise lesion symptom mappings on (i) the full sample (N=96) in which all patients with hand and/or finger imitation apraxia as well as without apraxia were included and (ii) three sub-samples, which excluded those patients from the full sample showing isolated hand imitation deficits, isolated finger imitation deficits or shared (finger and hand) imitation deficits. Anatomical analyses revealed a cortical dissociation of finger imitation deficits (located more anteriorly) and of hand imitation deficits (located more posteriorly). The presence of patients with shared deficits did indeed dilute associations that appeared stronger in the respective isolated samples. Also, brain regions truly associated with hand imitation deficits showed a positive bias for finger imitation deficits, when the sample contained patients with shared deficits. In addition, our frequentist parameters uncovered that some of our Bayesian evidence supported reverse associations (damage protecting from rather than increasing the deficit). Anatomo-behavioral analyses that analyse patients with shared (hand and finger) and isolated (hand or finger) imitation deficits together in one sample do indeed lead to undesirable biases. This explains why some earlier studies failed to detect the apparent neural dissociation between hand and finger imitation deficits.

AI-based prediction of academic success: Support for many, disadvantage for some?

The use of computational tools to predict academic success has become increasingly popular. Machine learning algorithms, trained on past study histories, have been shown to provide valid predictions. However, knowing about biases and unfairness in algorithms, one should take a closer look at these predictions. This paper explores the extent to which the predictive accuracy of academic success varies between specific groups of students, focusing on traditional and non-traditional students (NTS), who have not acquired a higher education entrance qualification at school. In a case study the study compares several popular algorithms and their prediction quality, and investigates whether misclassified NTS show positive or negative biases. Results revealed that the accuracy of predicting academic success for NTS was significantly lower than when considering all students as a whole. The direction of the distortion cannot be determined exactly due to small case numbers. The study emphasizes that the possibility of bias always has to be considered when predicting study success, and the use of such tools must ensure there are no undesirable biases that could affect certain students.

Moving beyond the Aerosol Climatology of WRF-Solar: A Case Study over the North China Plain

Abstract Numerical weather prediction (NWP), when accessible, is a crucial input to short-term solar power forecasting. WRF-Solar, the first NWP model specifically designed for solar energy applications, has shown promising predictive capability. Nevertheless, few attempts have been made to investigate its performance under high aerosol loading, which attenuates incoming radiation significantly. The North China Plain is a polluted region due to industrialization, which constitutes a proper testbed for such investigation. In this paper, aerosol direct radiative effect (DRE) on three surface shortwave radiation components (i.e., global, beam, and diffuse) during five heavy pollution episodes is studied within the WRF-Solar framework. Results show that WRF-Solar overestimates instantaneous beam radiation up to 795.3 W m−2 when the aerosol DRE is not considered. Although such overestimation can be partially offset by an underestimation of the diffuse radiation of about 194.5 W m−2, the overestimation of the global radiation still reaches 160.2 W m−2. This undesirable bias can be reduced when WRF-Solar is powered by Copernicus Atmosphere Monitoring Service (CAMS) aerosol forecasts, which then translates to accuracy improvements in photovoltaic (PV) power forecasts. This work also compares the forecast performance of the CAMS-powered WRF-Solar with that of the European Centre for Medium-Range Weather Forecasts model. Under high aerosol loading conditions, the irradiance forecast accuracy generated by WRF-Solar increased by 53.2% and the PV power forecast accuracy increased by 6.8%. Significance Statement Numerical weather prediction (NWP) is the “go-to” approach for achieving high-performance day-ahead solar power forecasting. Integrating time-varying aerosol forecasts into NWP models effectively captures aerosol direct radiation effects, thereby enhancing the accuracy of solar irradiance forecasts in heavily polluted regions. This work not only quantifies the aerosol effects on global, beam, and diffuse irradiance but also reveals the physical mechanisms of irradiance-to-power conversion by constructing a model chain. Using the North China Plain as a testbed, the performance of WRF-Solar on solar power forecasting on five severe pollution days is analyzed. This version of WRF-Solar can outperform the European Centre for Medium-Range Weather Forecasts model, confirming the need for generating high spatial–temporal NWP.

The Implementation of Recommender Systems for Mental Health Recovery Narratives: Evaluation of Use and Performance.

Recommender systems help narrow down a large range of items to a smaller, personalized set. NarraGive is a first-in-field hybrid recommender system for mental health recovery narratives, recommending narratives based on their content and narrator characteristics (using content-based filtering) and on narratives beneficially impacting other similar users (using collaborative filtering). NarraGive is integrated into the Narrative Experiences Online (NEON) intervention, a web application providing access to the NEON Collection of recovery narratives. This study aims to analyze the 3 recommender system algorithms used in NarraGive to inform future interventions using recommender systems for lived experience narratives. Using a recently published framework for evaluating recommender systems to structure the analysis, we compared the content-based filtering algorithm and collaborative filtering algorithms by evaluating the accuracy (how close the predicted ratings are to the true ratings), precision (the proportion of the recommended narratives that are relevant), diversity (how diverse the recommended narratives are), coverage (the proportion of all available narratives that can be recommended), and unfairness (whether the algorithms produce less accurate predictions for disadvantaged participants) across gender and ethnicity. We used data from all participants in 2 parallel-group, waitlist control clinical trials of the NEON intervention (NEON trial: N=739; NEON for other [eg, nonpsychosis] mental health problems [NEON-O] trial: N=1023). Both trials included people with self-reported mental health problems who had and had not used statutory mental health services. In addition, NEON trial participants had experienced self-reported psychosis in the previous 5 years. Our evaluation used a database of Likert-scale narrative ratings provided by trial participants in response to validated narrative feedback questions. Participants from the NEON and NEON-O trials provided 2288 and 1896 narrative ratings, respectively. Each rated narrative had a median of 3 ratings and 2 ratings, respectively. For the NEON trial, the content-based filtering algorithm performed better for coverage; the collaborative filtering algorithms performed better for accuracy, diversity, and unfairness across both gender and ethnicity; and neither algorithm performed better for precision. For the NEON-O trial, the content-based filtering algorithm did not perform better on any metric; the collaborative filtering algorithms performed better on accuracy and unfairness across both gender and ethnicity; and neither algorithm performed better for precision, diversity, or coverage. Clinical population may be associated with recommender system performance. Recommender systems are susceptible to a wide range of undesirable biases. Approaches to mitigating these include providing enough initial data for the recommender system (to prevent overfitting), ensuring that items can be accessed outside the recommender system (to prevent a feedback loop between accessed items and recommended items), and encouraging participants to provide feedback on every narrative they interact with (to prevent participants from only providing feedback when they have strong opinions).

Publication bias in psychology: A closer look at the correlation between sample size and effect size.

Previously observed negative correlations between sample size and effect size (n-ES correlation) in psychological research have been interpreted as evidence for publication bias and related undesirable biases. Here, we present two studies aimed at better understanding to what extent negative n-ES correlations reflect such biases or might be explained by unproblematic adjustments of sample size to expected effect sizes. In Study 1, we analysed n-ES correlations in 150 meta-analyses from cognitive, organizational, and social psychology and in 57 multiple replications, which are free from relevant biases. In Study 2, we used a random sample of 160 psychology papers to compare the n-ES correlation for effects that are central to these papers and effects selected at random from these papers. n-ES correlations proved inconspicuous in meta-analyses. In line with previous research, they do not suggest that publication bias and related biases have a strong impact on meta-analyses in psychology. A much higher n-ES correlation emerged for publications' focal effects. To what extent this should be attributed to publication bias and related biases remains unclear.

Biased Complementary-Label Learning Without True Labels.

In complementary-label learning (CLL), the complementary transition matrix, denoting the probabilities that true labels flip into complementary labels (CLs) which specify classes observations do not belong to, is crucial to building statistically consistent classifiers. Most existing works implicitly assume that the transition probabilities are identical, which is not true in practice and may lead to undesirable bias in solutions. Few recent works have extended the problem to a biased setting but limit their explorations to modeling the transition matrix by exploiting the complementary class posteriors of anchor points (i.e., instances that almost certainly belong to a specific class). However, due to the severe corruption and unevenness of biased CLs, both anchor points and complementary class posteriors are difficult to predict accurately in the absence of true labels. In this article, rather than directly predicting these two error-prone items, we instead propose a divided-T estimator as an alternative to effectively learn transition matrices from only biased CLs. Specifically, we exploit semantic clustering to mitigate the adverse effects arising from CLs. By introducing the learned semantic clusters as an intermediate class, we factorize the original transition matrix into the product of two easy-to-estimate matrices that are not reliant on the two error-prone items. Both theoretical analyses and empirical results justify the effectiveness of the divided- T estimator for estimating transition matrices under a mild assumption. Experimental results on benchmark datasets further demonstrate that the divided- T estimator outperforms state-of-the-art (SOTA) methods by a substantial margin.

Polyamines promote xenobiotic nucleic acid synthesis by modified thermophilic polymerase mutants.

The enzymatic synthesis of xenobiotic nucleic acids (XNA), which are artificially sugar-modified nucleic acids, is essential for the preparation of XNA libraries. XNA libraries are used in the in vitro selection of XNA aptamers and enzymes (XNAzymes). Efficient enzymatic synthesis of various XNAs can enable the screening of high-quality XNA aptamers and XNAzymes by expanding the diversity of XNA libraries and adding a variety of properties to XNA aptamers and XNAzymes. However, XNAs that form unstable duplexes with DNA, such as arabino nucleic acid (ANA), may dissociate during enzyme synthesis at temperatures suitable for thermophilic polymerases. Thus, such XNAs are not efficiently synthesised by the thermophilic polymerase mutants at the end of the sequence. This undesirable bias reduces the possibility of generating high-quality XNA aptamers and XNAzymes. Here, we demonstrate that polyamine-induced DNA/ANA duplex stabilisation promotes ANA synthesis that is catalysed by thermophilic polymerase mutants. Several polyamines, including spermine, spermidine, cadaverine, and putrescine promote ANA synthesis. The negative effect of polyamines on the fidelity of ANA synthesis was negligible. We also showed that polyamines promote the synthesis of other XNAs, including 2'-amino-RNA/2'-fluoro-RNA mixture and 2'-O-methyl-RNA. In addition, we found that polyamine promotes DNA synthesis from the 2'-O-methyl-RNA template. Polyamines, with the use of thermophilic polymerase mutants, may allow further development of XNA aptamers and XNAzymes by promoting the transcription and reverse transcription of XNAs.

Novel and flexible parameter estimation methods for data-consistent inversion in mechanistic modelling.

Predictions for physical systems often rely upon knowledge acquired from ensembles of entities, e.g. ensembles of cells in biological sciences. For qualitative and quantitative analysis, these ensembles are simulated with parametric families of mechanistic models (MMs). Two classes of methodologies, based on Bayesian inference and population of models, currently prevail in parameter estimation for physical systems. However, in Bayesian analysis, uninformative priors for MM parameters introduce undesirable bias. Here, we propose how to infer parameters within the framework of stochastic inverse problems (SIPs), also termed data-consistent inversion, wherein the prior targets only uncertainties that arise due to MM non-invertibility. To demonstrate, we introduce new methods to solve SIPs based on rejection sampling, Markov chain Monte Carlo, and generative adversarial networks (GANs). In addition, to overcome limitations of SIPs, we reformulate SIPs based on constrained optimization and present a novel GAN to solve the constrained optimization problem.

A graph-based probabilistic geometric deep learning framework with online enforcement of physical constraints to predict the criticality of defects in porous materials

Stress prediction in porous materials and structures is challenging due to the high computational cost associated with direct numerical simulations. Convolutional Neural Network (CNN) based architectures have recently been proposed as surrogates to approximate and extrapolate the solution of such multiscale simulations. These methodologies are usually limited to 2D problems due to the high computational cost of 3D voxel based CNNs. We propose a novel geometric learning approach based on a Graph Neural Network (GNN) that efficiently deals with three-dimensional problems by performing convolutions over 2D surfaces only. Following our previous developments using pixel-based CNN, we train the GNN to automatically add local fine-scale stress corrections to an inexpensively computed coarse stress prediction in the porous structure of interest. Our method is Bayesian and generates densities of stress fields, from which credible intervals may be extracted. As a second scientific contribution, we propose to improve the extrapolation ability of our network by deploying a strategy of online physics-based corrections. Specifically, we condition the posterior predictions of our probabilistic predictions to satisfy partial equilibrium at the microscale, at the inference stage. This is done using an Ensemble Kalman algorithm, to ensure tractability of the Bayesian conditioning operation. We show that this innovative methodology allows us to alleviate the effect of undesirable biases observed in the outputs of the uncorrected GNN, and improves the accuracy of the predictions in general.

Poster] Classification of Holdings in Flights Arriving at Dubai International Airport (DXB) in One Year

The purpose of this study was to use a machine-learning approach for the classification of holdings for all flights arriving at Dubai International Airport (IATA DXB) in a period of one year. The study used data from ADS-B for all flights arriving and departing DXB for the period 15 February 2018 and 15 February 2019. For all 189,999 arrivals analyzed, it was identified the Standard Terminal Arrival Route (STAR) flown, the occurrence of holdings, the aircraft flying (for the determination of wake turbulence category), the visibility, and data related to the interception of a circle centered around DXB with a radius of 50 NM (geographical position, number of aircraft entering the circle for several time windows, number of aircraft flying within this circle etc), along with the number of aircraft taking-off and landing during a given time period. These features were tested, in different combinations, to determine their impact on metrics used to evaluate the classifier output quality. After running classifiers designed with different algorithms and combinations of features, it was identified the one which, with the CatBoost algorithm, gave the best F1 (0.777), Precision (0.847) and Accuracy (0.912) for a period of 10 minutes after interception. The F1 increased to 0.85 when the altitude and the speed at the interception of the circle were included as features, but they were discarded as they introduced an undesirable bias in the final model.

Simultaneous selection and incorporation of consistent external aggregate information.

Interest has grown in synthesizing participant level data of a study with relevant external aggregate information. Several efficient and flexible procedures have been developed under the assumption that the internal study and the external sources concern the same population. This homogeneity condition, albeit commonly being imposed, is hard to check due to limitedly available external information in aggregate data forms. Bias may be introduced when the assumption is violated. In this article, we propose a penalized likelihood approach that avoids undesirable bias by simultaneously selecting and synthesizing consistent external aggregate information. The proposed approach provides a general framework which incorporate consistent external information from heterogeneous study populations as long as the conditional distribution of the dependent variable under investigation is same and differences in the independent variable distributions are properly accounted for via a semi-parametric density ratio model. The proposed approach also properly accounts for the sampling errors in the external information. A two-step estimator and an optimization algorithm are proposed for computation. We establish the selection and estimation consistency and the asymptotic normality of the two-step estimator. The proposed approach is illustrated with an analysis of gestational weight gain management studies.

Testing transpiration rates of juvenile Aleppo pine trees using the heat ratio method under laboratory conditions

AbstractTree transpiration considerably contributes to evaporative fluxes to the atmosphere in terrestrial ecosystems. Accurate transpiration quantification provides relevant information about forest water use and may benefit adaptive forest management, especially in a global change context. Tree transpiration can be measured by several methods, and sap flow measurements are one of the most valued. However, species‐specific validations of these techniques are required to avoid undesirable bias. This is especially relevant in species with low transpiration rates where errors may be relevant, such as Aleppo pine trees (Pinus halepensis Mill.). Moreover, another significant source of uncertainty in sap flow measurements is probe misalignment. Hence, the aim of this study was to correlate transpiration rates estimated by sap flow probes using the heat ratio method (THRM) and load cells to independently monitor water transpiration in juvenile Aleppo pine trees. The corrections to improve transpiration measures, including misalignment correction, were applied to THRM results to test if the accuracy of results improved. These measurements were recorded in greenhouse under controlled conditions to implement different environmental conditions. The environmental variables that ruled the experiment, mainly vapour pressure deficit and soil water availability, spanned in a wide range of values. The results showed an accurate linear correspondence between TOBS and THRM for low and medium values, but moderate underestimations at high transpiration rates were observed. These underestimations were partly removed when applying probe misalignment correction. This study supports the notion that HRM offers accurate Aleppo pine transpiration estimations with low and medium values under a variety of abiotic conditions, which also has implications for HRM application in other isohydric species. The results also support the interest in the use of probe misalignment correction to estimate transpiration, mainly when high transpiration values are recorded. The results of this study can be considered as a preliminary approach for future research in order to improve the estimates of the transpiration rates of the Aleppo pine under the limiting conditions of the Mediterranean.

Characterization of the Global Bias Problem in Aerial Federated Learning

Unmanned aerial vehicles (UAVs) mobility enables flexible and customized federated learning (FL) at the network edge. However, the underlying uncertainties in the aerial-terrestrial wireless channel may lead to a biased FL model. In particular, the distribution of the global model and the aggregation of the local updates within the FL learning rounds at the UAVs are governed by the reliability of the wireless channel. This creates an undesirable bias towards the training data of ground devices with better channel conditions, and vice versa. This paper characterizes the global bias problem of aerial FL in large-scale UAV networks. To this end, the paper proposes a channel-aware distribution and aggregation scheme to enforce equal contribution from all devices in the FL training as a means to resolve the global bias problem. We demonstrate the convergence of the proposed method by experimenting with the MNIST dataset and show its superiority compared to existing methods. The obtained results enable system parameter tuning to relieve the impact of the aerial channel deficiency on the FL convergence rate.

A Review Paper: Will Artificial Intelligence (AI) Replace the Human Recruiter?

AI’s influence on the process of recruitment is continually showing upward growth. From screening and selecting the best-fit candidates using algorithms, it is now assisting people with analytics to develop human resources and even designing models to predict the salary of the newcomer. Despite this ever expanding and all- encompassing role, it is being criticised for bringing forth algorithm related biases, which it is supposed to reduce/ eliminate. AI’s advanced offering ChatGPT also threatens to jeopardise the existence of the recruiter. However, the human element (particularly, the recruiter) is needed to monitor the entire algorithm process to eliminate the undesired biases and to prevent new ones. The intuitive nuance and critical thinking ability of the recruiter cannot be dispensed with. The recruiter also needs to be compassionate and only a human recruiter can fit into this bill. A human recruiter is also needed to recognise, appreciate and value a diversified workforce, thereby looking beyond the mere binary focus of machines. A balance between subjective (human) and objective (AI) ensuring “positive-sum automation” will always be needed for business decision- making, including the process of recruitment.

#5490 GENERATIVE ARTIFICIAL INTELLIGENCE FOR CREATION OF SYNTHETIC HYPERTENSION TRIAL DATA

Abstract Background and Aims Synthetic data can be an effective supplement or alternative to real data for the training of machine learning models. Synthetic data may also be used to evaluate new tools, develop educational curricula, or remove undesirable biases in datasets. We aim to evaluate four synthetic data generation methods applied to hypertension randomized clinical trial data. Method The Systolic Blood Pressure Intervention Trial (SPRINT) trial showed that intensive BP control to SBP <120 mm Hg results in significant cardiovascular benefits in high-risk patients with hypertension compared with routine BP control to <140 mm Hg. The Synthetic Data Vault (SDV) is a Synthetic Data Generation ecosystem of libraries that allows users to easily generate new Synthetic Data that has the same format and statistical properties as the original dataset. SDV supports multiple types of data, including date-times, discrete-ordinal, categorical, and numerical. SPRINT data was pre-processed to create a single table of 140,000 patient visits with baseline variables (age, sex, race, aspirin use, estimated Glomerular Filtration Rate (eGFR)) and visit level variables (systolic and diastolic blood pressure, heart rate and total number of antihypertensive medications at end of visit). Using the SDV library for python, we used four generative models to create synthetic SPRINT data, 1. Gaussian copula model, 2. Conditional Tabular Generative adversarial network (CTGAN), 3. CopulaGan model, and 4. Tabular Variational Auto-encode (TVAE). We evaluated the results using the SDMetrics library which includes the shapes of the columns (marginal distributions), the pairwise trends between the columns (correlations), reproduce mathematical properties from your original data and new row synthesis. Finally, an overall quality score which represents an amalgamation of the marginal distribution and correlations was computed, where 0 indicates the lowest quality and 1 indicates the highest. Results Two hundred thousand synthetic patient visits were created for each method. The overall quality scores in order were 90.67% for Gaussian copula, 86.77% for TVAE, 81.03% for CTGAN’, and 79.7% for CopulaGAN. The column shape score which represents the marginal distribution was highest for Gaussian Copula (94.54%), followed by TVAE (88.44%), CTGAN (82.35%), and Copula GAN (80.27%). The column pair trend which corresponds to correlations was highest for Gaussian Copula (86.8%), followed by TAVE (85.1%), CTGAN (79.72%), and Copula GAN (79.12%). Conclusion Gaussian copula created the highest scoring synthetic SPRINT data based on the marginal distribution, correlations, and overall score. The Synthetic Data Vault is a feasible collection of methods for generation of synthetic clinical trial data for training future machine learning and AI models.

Effects of the Vertical Component of Ground Motion on the Maximum-Inelastic-Horizontal-Response of SDOFs

This study investigates potential effects of the vertical-component-of-ground-motion (VCGM) on the maximum-inelastic-horizontal-response (MIHR) of single-degree-of-freedom (SDOF) systems. The effects of VCGM can be considered in two categories. First, it may initiate some modes of failure, like over-compression, and second, it might adversely influence the horizontal response through altering the P-delta effect. The latter is a result of variation of the gravitational acceleration, and hence, seismic weight, upon action of the VCGM. This study uses time-history analysis to compare the response of SDOFs in presence and absence of the VCGM. In both cases, the P-delta effect is inherent in analyses and period-dependent feature of the stability-coefficient, which is essential for reliable treatment of the P-delta effect, is explicitly reflected. Since available record selection and scaling strategies use uncoupled vibrators in vertical and horizontal directions, this research follows two steps to avoid undesirable bias. The first step, which focuses on examining the influence of some important parameters, adopts an event-based record selection scheme. Results indicate that for certain combinations of the strength reduction factor and initial period of vibration the effect of the VCGM on the MIHR is significant. Specifically, the systems with initial periods between 1.0 and 2.0 s are found most vulnerable. Moreover, the mentioned effect is not limited to near source regions. The second step evaluates capability of the VCGM to initiate dynamic instability. In this step, a set of 26 records, which resembles an assumed target spectrum at a near field site, is used. Results show that under action of some of the examined records, the presence of the VCGM leads to dynamic instability; hence, this study suggests consideration of the VCGM for collapse evaluation.

Bottlenecks CLUB: Unifying Information-Theoretic Trade-Offs Among Complexity, Leakage, and Utility

Bottleneck problems are an important class of optimization problems that have recently gained increasing attention in the domain of machine learning and information theory. They are widely used in generative models, fair machine learning algorithms, design of privacy-assuring mechanisms, and appear as information-theoretic performance bounds in various multi-user communication problems. In this work, we propose a general family of optimization problems, termed as <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">complexity-leakage-utility bottleneck (CLUB)</i> model, which (i) provides a unified theoretical framework that generalizes most of the state-of-the-art literature for the information-theoretic privacy models, (ii) establishes a new interpretation of the popular generative and discriminative models, (iii) constructs new insights for the generative compression models, and (iv) can be used to obtain fair generative models. We first formulate the CLUB model as a complexity-constrained privacy-utility optimization problem. We then connect it with the closely related bottleneck problems, namely information bottleneck (IB), privacy funnel (PF), deterministic IB (DIB), conditional entropy bottleneck (CEB), and conditional PF (CPF). We show that the CLUB model generalizes all these problems as well as most other information-theoretic privacy models. Then, we construct the deep variational CLUB (DVCLUB) models by employing neural networks to parameterize variational approximations of the associated information quantities. Building upon these information quantities, we present unified objectives of the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">supervised</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">unsupervised</i> DVCLUB models. Leveraging the DVCLUB model in an unsupervised setup, we then connect it with state-of-the-art generative models, such as variational auto-encoders (VAEs), generative adversarial networks (GANs), as well as the Wasserstein GAN (WGAN), Wasserstein auto-encoder (WAE), and adversarial auto-encoder (AAE) models through the optimal transport (OT) problem. We then show that the DVCLUB model can also be used in fair representation learning problems, where the goal is to mitigate the undesired bias during the training phase of a machine learning model. We conduct extensive quantitative experiments on colored-MNIST and CelebA datasets.

Interclass Similarity Transfer for Imbalanced Aerial Scene Classification

Imbalanced class distributions widely exist in real-world aerial images, which brings a significant challenge to aerial scene classification due to the undesirable bias towards the majority classes as well as overfitting for the minority classes. Although the similarity between different scene classes may be inconsistent, they can be measured by the mean of feature statistics. This motivates us to transfer the statistics of the majority class to the minority class having similar feature statistics. Specifically, based on the observation that the feature statistic of each class may follow the Gaussian distribution, the similarity across different classes would thus be described by the mean of feature statistics. The distributions of minority classes would afterward be calibrated by statistical transfer via interclass similarity, and a sufficient number of features could hence be generated for the minority class to improve its performance in classifier learning. We demonstrate the effectiveness of the proposed method for imbalanced aerial scene classification on the imbalanced AID and NWPU-RESISC45 datasets. The proposed method outperforms alternatives by a large margin in both overall performance and minority classification performance of imbalanced aerial scenes.

ComBat Harmonization: Empirical Bayes versus fully Bayes approaches

Studying small effects or subtle neuroanatomical variation requires large-scale sample size data. As a result, combining neuroimaging data from multiple datasets is necessary. Variation in acquisition protocols, magnetic field strength, scanner build, and many other non-biologically related factors can introduce undesirable bias into studies. Hence, harmonization is required to remove the bias-inducing factors from the data. ComBat is one of the most common methods applied to features from structural images. ComBat models the data using a hierarchical Bayesian model and uses the empirical Bayes approach to infer the distribution of the unknown factors. The empirical Bayes harmonization method is computationally efficient and provides valid point estimates. However, it tends to underestimate uncertainty. This paper investigates a new approach, fully Bayesian ComBat, where Monte Carlo sampling is used for statistical inference. When comparing fully Bayesian and empirical Bayesian ComBat, we found Empirical Bayesian ComBat more effectively removed scanner strength information and was much more computationally efficient. Conversely, fully Bayesian ComBat better preserved biological disease and age-related information while performing more accurate harmonization on traveling subjects. The fully Bayesian approach generates a rich posterior distribution, which is useful for generating simulated imaging features for improving classifier performance in a limited data setting. We show the generative capacity of our model for augmenting and improving the detection of patients with Alzheimer’s disease. Posterior distributions for harmonized imaging measures can also be used for brain-wide uncertainty comparison and more principled downstream statistical analysis.Code for our new fully Bayesian ComBat extension is available at https://github.com/batmanlab/BayesComBat.