Search

Recommender systems are ubiquitous on various online platforms and provide significant value to the users in helping them find relevant content/items to consume. After item consumption, users can often provide feedback (i.e., their preference ratings for the item) to the system. Research studies have shown that recommender systems' predictions, observed by users, can cause biases in users' postconsumption preference ratings. Because these ratings are typically fed back to the system as training data for future predictions, this process is likely to influence the system's performance over time. We use a simulation approach to investigate the longitudinal impact of preference biases on the dynamics of recommender systems' performance. Our results reveal that preference biases significantly impair recommendation performance and users' consumption outcomes, and larger biases cause disproportionately large negative effects. Additionally, less popular and less distinctive (in terms of their content) items are more susceptible to preference biases. Furthermore, considering the substantial impact of preference biases on recommendation performance, we examine the issue of debiasing user-submitted ratings. We find that relying solely on historical rating data is unlikely to be effective in debiasing; thus, we propose/evaluate new debiasing approaches that use additional relevant information that can be collected by recommendation platforms. Research studies have shown that recommender systems' predictions that are observed by users can cause biases in users' postconsumption preference ratings. This can happen as part of the standard, normal system use, where biases are typically caused by the system's inherent prediction errors (i.e., because of the less-than-perfect accuracy of recommendation methods). Because users' preference ratings are typically fed back to the system as training data for future predictions, this process is likely to influence the performance of the system in the long run. We use a simulation approach to study the longitudinal impact of preference biases (and their magnitude) on the dynamics of recommender systems' performance. Our simulation results show that preference biases significantly impair the system's prediction performance (i.e., prediction accuracy) as well as users' consumption outcomes (i.e., consumption relevance and diversity) over time. The impact is nonlinear to the size of the bias, that is, large bias causes disproportionately large negative effects. Also, items that are less popular and less distinctive (in terms of their content) are affected more by preference biases. Furthermore, given the impact of preference bias on the recommender systems' performance, we explore the problem of debiasing user-submitted ratings. We empirically demonstrate that relying solely on historical rating data is unlikely to be effective in debiasing. We also propose and evaluate two debiasing approaches that take into account additional relevant information that can be collected by recommendation platforms. Our findings provide important implications for the design of recommender systems. History: Olivia Liu Sheng, Senior Editor; Huimin Zhao, Associate Editor. Supplemental Material: The e-companion is available at https://doi.org/10.1287/isre.2021.0133. [ABSTRACT FROM AUTHOR]

A typical Monte Carlo rendering method contributes one light path only to a single pixel at a time. Reusing light paths across multiple pixels, however, can amortize the cost and improve the efficiency. The state of the art of path reuse is to employ shift mapping to reduce the cost of path reuse, while its computation cost is still proportional to the number of pixels processed in shift mapping. We propose a general framework for efficiently reusing light paths to multiple pixels arranged in arbitrary two-dimensional shapes. Our shape is defined as a set of multiple pixels, and the framework allows us to reuse light paths among pixels in a shape faster than simply evaluating all pixels via shift mapping. The key idea is to sparsely evaluate the contribution of shifted paths at random pixels within the shape and interpolate the contribution to the other pixels. We apply a debiasing estimator to ensure unbiasedness. Our method can be integrated with many existing rendering methods and brings consistent improvement over its single-pixel counterpart. [ABSTRACT FROM AUTHOR]

We consider the problem of estimating a low-dimensional parameter in high-dimensional linear regression. Constructing an approximately unbiased estimate of the parameter of interest is a crucial step towards performing statistical inference. Several authors suggest to orthogonalize both the variable of interest and the outcome with respect to the nuisance variables, and then regress the residual outcome with respect to the residual variable. This is possible if the covariance structure of the regressors is perfectly known, or is sufficiently structured that it can be estimated accurately from data (e.g. the precision matrix is sufficiently sparse). Here we consider a regime in which the covariate model can only be estimated inaccurately, and hence existing debiasing approaches are not guaranteed to work. We propose the correlation adjusted debiased Lasso , which nearly eliminates this bias in some cases, including cases in which the estimation errors are neither negligible nor orthogonal. [ABSTRACT FROM AUTHOR]

Decision-making processes are vulnerable to cognitive biases like hindsight bias, with particularly fateful consequences in forensic contexts. However, while debiasing strategies have been effective in various areas, their impact in forensics is underexplored. We investigated hindsight bias and a simple awareness-based debiasing strategy in novices (n = 52) and forensic professionals (n = 49). Participants were assigned to baseline, biased, or debiased conditions and rated an offender’s risk of re-offending using case vignettes. Significant hindsight bias was found in novices, but not experts who were also more aware of biases. Debiasing proved effective in novices, indicating that raising awareness may enhance equitable forensic decision-making. [ABSTRACT FROM AUTHOR]

To demonstrate the feasibility of estimating a social tariff free of utility curvature and probability weighting biases and to test transferability between riskless and risky contexts. Valuations for a selection of EQ-5D-3L health states were collected from a large and representative sample (N = 1676) of the Spanish general population through computer-assisted personal interviewing. Two elicitation methods were used: the traditional time trade-off (TTO) and a novel risky-TTO procedure. Both methods are equivalent for better than death states, which allowed us to test transferability of utilities across riskless and risky contexts. Corrective procedures applied are based on rank-dependent utility theory, identifying parameter estimates at the individual level. All corrections are health-state specific, which is a unique feature of our corrective approach. Two corrected value sets for the EQ-5D-3L system are estimated, highlighting the feasibility of developing national tariffs under nonexpected utility theories, such as rank-dependent utility. Furthermore, transferability was not supported for at least half of the health states valued by our sample. It is feasible to estimate a social tariff by using interviewing techniques, sample sizes, and sample representativeness equivalent to prior studies designed to generate national value sets for the EQ-5D. Utilities obtained in distinct contexts may not be interchangeable. Our findings caution against routinely taking transferability of utility for granted. • To date there is no national social tariff or value set fully corrected under nonexpected utility assumptions. • This article presents 2 value sets for the EQ-5D-3L system estimated by applying corrections based on rank-dependent utility theory, identifying parameter estimates at individual level. All corrections done are health-state specific, which is a unique feature of our corrective approach. These findings highlight the feasibility of developing national tariffs under nonexpected utility theories. • Our results suggest that utilities obtained in distinct contexts may not be interchangeable. Transferability across riskless and risky domains was not supported for at least half of the health states valued by our sample. Consequently, it seems that ensuring tariffs are not context dependent is as relevant and challenging as correcting value sets. [ABSTRACT FROM AUTHOR]

We use the situational theory of problem solving to explicate how publics engage in conspiratorial thinking as a form of cognitive problem solving. In doing so, we develop a new typology of conspiracy theories and introduce conceptual definitions and operationalizations of conspiratorial thinking as both dispositional and situational. We conduct two survey studies which provide evidence of the measures' reliability and validity. We also investigate the debiasing effects of relational and informational strategies employed by social institutions. Finally, we present institutional behavioral strategies, grounded in the strategic behavioral theory of communication management, that can reduce conspiratorial thinking. Our new concepts and measurement approaches describe the origins and processes of public conspiracism and provide a foundation for implementing ethical and effective interventions, ultimately contributing to a more informed and educated society. [ABSTRACT FROM AUTHOR]

Nascent entrepreneurs often believe that their chances of success are better than those of others due to imperfect information about the competencies and accomplishments of other entrepreneurs, leading to overplacement. Theory suggests that the provision of historical outcome data of comparable projects could help entrepreneurs develop more realistic plans and expectations by closing the information gap and enabling the calibration of their beliefs. However, effectively calibrating beliefs by incorporating new reference information requires effortful cognitive processing and rational integration of the data, which may be impeded by the same cognitive biases leading to overplacement initially. Drawing from a unique dataset that allows us to observe substantial parts of the planning process of 971 entrepreneurs, we investigate the effectiveness of providing reference values as a debiasing tool. Rather than rationally leveraging the information for honest self-assessment, our findings suggest that entrepreneurs use the information to differentiate themselves even more from the reference group after they see the historical values. This, in turn, results in an even higher level of overplacement. JEL Classifications: G41, D91, L26, L25. [ABSTRACT FROM AUTHOR]

With the proliferation of social networks and their impact on human life, one of the rising problems in this environment is the rise in verbal and written insults and hatred. As one of the significant platforms for distributing text-based content, Twitter frequently publishes its users' abusive remarks. Creating a model that requires a complete collection of offensive sentences is the initial stage in recognizing objectionable phrases. In addition, despite the abundance of resources in English and other languages, there are limited resources and studies on identifying hateful and offensive statements in Persian. In this study, we compiled a 38K-tweet dataset of Persian Hate and Offensive language using keyword-based data selection strategies. A Persian offensive lexicon and nine hatred target group lexicons were gathered through crowdsourcing for this purpose. The dataset was annotated manually so that at least two annotators investigated tweets. In addition, for the purpose of analyzing the effect of used lexicons on language model functionality, we employed two assessment criteria (FPED and pAUCED) to measure the dataset's potential bias. Then, by configuring the dataset based on the results of the bias measurement, we mitigated the effect of words' bias in tweets on language model performance. The results indicate that bias is significantly diminished, while less than a hundredth reduced the F1 score. [ABSTRACT FROM AUTHOR]

Algorithmic bias occurs when algorithms incorporate biases in the human decisions on which they are trained. We find that people see more of their biases (e.g., age, gender, race) in the decisions of algorithms than in their own decisions. Research participants saw more bias in the decisions of algorithms trained on their decisions than in their own decisions, even when those decisions were the same and participants were incentivized to reveal their true beliefs. By contrast, participants saw as much bias in the decisions of algorithms trained on their decisions as in the decisions of other participants and algorithms trained on the decisions of other participants. Cognitive psychological processes and motivated reasoning help explain why people see more of their biases in algorithms. Research participants most susceptible to bias blind spot were most likely to see more bias in algorithms than self. Participants were also more likely to perceive algorithms than themselves to have been influenced by irrelevant biasing attributes (e.g., race) but not by relevant attributes (e.g., user reviews). Because participants saw more of their biases in algorithms than themselves, they were more likely to make debiasing corrections to decisions attributed to an algorithm than to themselves. Our findings show that bias is more readily perceived in algorithms than in self and suggest how to use algorithms to reveal and correct biased human decisions. [ABSTRACT FROM AUTHOR]

Causal illusions consist of believing that there is a causal relationship between events that are actually unrelated. This bias is associated with pseudoscience, stereotypes and other unjustified beliefs. Thus, it seems important to develop educational interventions to reduce them. To our knowledge, the only debiasing intervention designed to be used at schools was developed by Barberia et al. (Barberia et al. 2013 PLoS One 8, e71303 (doi:10.1371/journal.pone.0071303)), focusing on base rates, control conditions and confounding variables. Their assessment used an active causal illusion task where participants could manipulate the candidate cause. The intervention reduced causal illusions in adolescents but was only tested in a small experimental project. The present research evaluated it in a large-scale project through a collaboration with the Spanish Foundation for Science and Technology (FECYT), and was conducted in schools to make it ecologically valid. It included a pilot study (n = 287), a large-scale implementation (n = 1668; 40 schools) and a six-month follow-up (n = 353). Results showed medium-to-large and long-lasting effects on the reduction of causal illusions. To our knowledge, this is the first research showing the efficacy and long-term effects of a debiasing intervention against causal illusions that can be used on a large scale through the educational system.

Due to the rapidly changing climate, the frequency and severity of extreme weather is expected to increase over the coming decades. As fully‐resolved climate simulations remain computationally intractable, policy makers must rely on coarse‐models to quantify risk for extremes. However, coarse models suffer from inherent bias due to the ignored "sub‐grid" scales. We propose a framework to non‐intrusively debias coarse‐resolution climate predictions using neural‐network (NN) correction operators. Previous efforts have attempted to train such operators using loss functions that match statistics. However, this approach falls short with events that have longer return period than that of the training data, since the reference statistics have not converged. Here, the scope is to formulate a learning method that allows for correction of dynamics and quantification of extreme events with longer return period than the training data. The key obstacle is the chaotic nature of the underlying dynamics. To overcome this challenge, we introduce a dynamical systems approach where the correction operator is trained using reference data and a coarse model simulation nudged toward that reference. The method is demonstrated on debiasing an under‐resolved quasi‐geostrophic model and the Energy Exascale Earth System Model (E3SM). For the former, our method enables the quantification of events that have return period two orders longer than the training data. For the latter, when trained on 8 years of ERA5 data, our approach is able to correct the coarse E3SM output to closely reflect the 36‐year ERA5 statistics for all prognostic variables and significantly reduce their spatial biases. Plain Language Summary: We present a general framework to design machine learned correction operators to improve the predicted statistics of low‐resolution climate simulations. We illustrate the approach, which acts on existing data in a post‐processing manner, on a simplified prototype climate model as well as a realistic climate model, namely the Energy Exascale Earth System Model (E3SM) with 110 km resolution. For the latter, we show that the developed approach is able to correct the low‐resolution E3SM output to closely reflect the climate statistics of historical observations as quantified by the ERA5 data set. We also demonstrate that our model significantly improves the prediction of atmospheric rivers, an example of extreme weather events resolvable by the low resolution model. Key Points: Development of non‐intrusive correction operators for coarse scale climate simulationsDesign of a training procedure that improves dynamics and allows for characterization of extremes with return period longer than the training dataApplication to Energy Exascale Earth System Model and demonstration of improvement on global and regional statistics [ABSTRACT FROM AUTHOR]

In experimental work, researchers have found that policymakers discount the opinions of constituents with whom they disagree. We build on these results with a national sample of local policymakers in the United States, exploring whether communicators can prevent policymakers from discounting their opinions by providing evidence of their own knowledge about a topic. We find that policymakers discount the opinions of hypothetical constituents with whom they disagree, but there is evidence that providing unambiguous evidence about a letter-writer's positive traits can reduce this discounting. We conclude with a discussion of implications for theory as well as practical implications for communicating with policymakers. [ABSTRACT FROM AUTHOR]

Copyright of Journal for Technology in Theory & Practice / Zeitschrift für Technikfolgenabschätzung in Theorie und Praxis (TATuP) is the property of Oekom Verlag GmbH and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Recent debiasing studies have shown that a short, plain-English explanation of the correct solution strategy can improve reasoning performance. However, these studies have predominantly focused on English-speaking populations, who were tested with problem contents designed for an English-speaking test environment. Here we explore whether the key findings of previous debiasing studies can be extended to native French speakers living in continental Europe (France). We ran a training session with a battery of three reasoning tasks (i.e., base-rate neglect, conjunction fallacy, and bat-and-ball) on 147 native French speakers. We used a two-response paradigm in which participants first gave an initial intuitive response, under time pressure and cognitive load, and then gave a final response after deliberation. Results showed a clear training effect, as early as the initial (intuitive) stage. Immediately after training, most participants solved the problems correctly, without the need for a deliberation process. The findings confirm that the intuitive debiasing training effect extends to native French speakers.

Abstract Due to the rapidly changing climate, the frequency and severity of extreme weather is expected to increase over the coming decades. As fully‐resolved climate simulations remain computationally intractable, policy makers must rely on coarse‐models to quantify risk for extremes. However, coarse models suffer from inherent bias due to the ignored “sub‐grid” scales. We propose a framework to non‐intrusively debias coarse‐resolution climate predictions using neural‐network (NN) correction operators. Previous efforts have attempted to train such operators using loss functions that match statistics. However, this approach falls short with events that have longer return period than that of the training data, since the reference statistics have not converged. Here, the scope is to formulate a learning method that allows for correction of dynamics and quantification of extreme events with longer return period than the training data. The key obstacle is the chaotic nature of the underlying dynamics. To overcome this challenge, we introduce a dynamical systems approach where the correction operator is trained using reference data and a coarse model simulation nudged toward that reference. The method is demonstrated on debiasing an under‐resolved quasi‐geostrophic model and the Energy Exascale Earth System Model (E3SM). For the former, our method enables the quantification of events that have return period two orders longer than the training data. For the latter, when trained on 8 years of ERA5 data, our approach is able to correct the coarse E3SM output to closely reflect the 36‐year ERA5 statistics for all prognostic variables and significantly reduce their spatial biases.

As arbiters of law and fact, judges are supposed to decide cases impartially, basing their decisions on authoritative legal sources and not being influenced by irrelevant factors. Empirical evidence, however, shows that judges are often influenced by implicit biases, which can affect the impartiality of their judgment and pose a threat to the right to a fair trial. In recent years, artificial intelligence (AI) has been increasingly used for a variety of applications in the public domain, often with the promise of being more accurate and objective than biased human decision-makers. Given this backdrop, this research article identifies how AI is being deployed by courts, mainly as decision-support tools for judges. It assesses the potential and limitations of these tools, focusing on their use for risk assessment. Further, the article shows how AI can be used as a debiasing tool, i. e., to detect patterns of bias in judicial decisions, allowing for corrective measures to be taken. Finally, it assesses the mechanisms and benefits of such use.

In this work, we tackle class incremental learning (CIL) for video action recognition, a relatively under-explored problem despite its practical importance. Directly applying image-based CIL methods does not work well in the video action recognition setting. We hypothesize the major reason is the spurious correlation between the action and background in video action recognition datasets/models. Recent literature shows that the spurious correlation hampers the generalization of models in the conventional action recognition setting. The problem is even more severe in the CIL setting due to the limited exemplars available in the rehearsal memory. We empirically show that mitigating the spurious correlation between the action and background is crucial to the CIL for video action recognition. We propose to learn background invariant action representations in the CIL setting by providing training videos with diverse backgrounds generated from background augmentation techniques. We validate the proposed method on public benchmarks: HMDB-51, UCF-101, and Something-Something-v2. • Video-based class incremental learning (video CIL) is important, yet under-explored. • Identify the background bias problem in video CIL setting. • The proposed method diversifies the scenes of training videos in the CIL setting. • The proposed method is a plug-and-play video CIL method. • The proposed method is tested UCF-101, HMDB-51, and Something-Something v2. [ABSTRACT FROM AUTHOR]

Causal illusions consist of believing that there is a causal relationship between events that are actually unrelated. This bias is associated with pseudoscience, stereotypes and other unjustified beliefs. Thus, it seems important to develop educational interventions to reduce them. To our knowledge, the only debiasing intervention designed to be used at schools was developed by Barberia et al. (Barberia et al. 2013 PLoS One 8 , e71303 (doi:10.1371/journal.pone.0071303)), focusing on base rates, control conditions and confounding variables. Their assessment used an active causal illusion task where participants could manipulate the candidate cause. The intervention reduced causal illusions in adolescents but was only tested in a small experimental project. The present research evaluated it in a large-scale project through a collaboration with the Spanish Foundation for Science and Technology (FECYT), and was conducted in schools to make it ecologically valid. It included a pilot study ( n = 287), a large-scale implementation ( n = 353). Results showed medium-to-large and long-lasting effects on the reduction of causal illusions. To our knowledge, this is the first research showing the efficacy and long-term effects of a debiasing intervention against causal illusions that can be used on a large scale through the educational system.n = 353). Results showed medium-to-large and long-lasting effects on the reduction of causal illusions. To our knowledge, this is the first research showing the efficacy and long-term effects of a debiasing intervention against causal illusions that can be used on a large scale through the educational system., Competing Interests: We declare we have no competing interests., (© 2024 The Author(s).)

An important problem in compositional data analysis is variable selection in linear regression models with compositional covariates. In the context of microbiome data analysis, there is a demand for considering grouping information such as structures among taxa and multiple sampling sites, resulting in multiple compositional covariates. We develop and compare two different methods of variable selection and inference strategies, based on the debiased lasso and a resampling-based approach. Confidence intervals for individual regression coefficients, obtained from each of the two methods, are shown to be asymptotically valid even in a high-dimension, low-sample-size regime. However, microbiome data often have extremely small sample sizes, rendering asymptotic results unreliable. Through extensive numerical comparisons of the finite-sample performances of the two methods, we find that resampling-based approaches outperform the debiased compositional lasso in cases of extremely small sample sizes, showing higher positive predictive values. Conversely, for larger sample sizes, debiasing yields better results. We apply the proposed multiple compositional regression to steer microbiome data, identifying key bacterial taxa associated with important cattle quality measures. • Resampling-based inference tools for multiple compositional regressions developed • Valid CIs obtained using debiased lasso and resampling • Performances of statistical inference compared in the case of low sample size [ABSTRACT FROM AUTHOR]

Understanding and reducing intergroup discrimination is at the forefront of psychological research. However, efforts to find flexible, scalable, and durable interventions to reduce discrimination have produced only mixed results. In this review, we highlight one potential avenue for developing new strategies for addressing discrimination: adapting prior research on debiasing—the process of lessening bias in judgment errors (e.g., motivated reasoning, overconfidence, and the anchoring heuristic). We first introduce a taxonomy for understanding intervention strategies that are common in the debiasing literature, then highlight existing approaches that have already proven successful for decreasing intergroup discrimination. Finally, we draw attention to promising debiasing interventions that have not yet been applied to the context of discrimination. A greater understanding of prior efforts to mitigate judgment biases more generally can expand efforts to reduce discrimination.Scientists studying intergroup biases are often concerned with lessening discrimination (unequal treatment of one social group versus another), but many interventions for reducing such biased behavior have weak or limited evidence. In this review article, we argue one productive avenue for reducing discrimination comes from adapting interventions in a separate field—judgment and decision-making—that has historically studied “debiasing”: the ways people can lessen the unwanted influence of irrelevant information on decision-making. While debiasing research shares several commonalities with research on reducing intergroup discrimination, many debiasing interventions have relied on methods that differ from those deployed in the intergroup bias literature. We review several instances where debiasing principles have been successfully applied toward reducing intergroup biases in behavior and introduce other debiasing techniques that may be well-suited for future efforts in lessening discrimination. [ABSTRACT FROM AUTHOR]

Recent debiasing studies have shown that a short, plain-English explanation of the correct solution strategy can improve reasoning performance. However, these studies have predominantly focused on English-speaking populations, who were tested with problem contents designed for an English-speaking test environment. Here we explore whether the key findings of previous debiasing studies can be extended to native French speakers living in continental Europe (France). We ran a training session with a battery of three reasoning tasks (i.e., base-rate neglect, conjunction fallacy, and bat-and-ball) on 147 native French speakers. We used a two-response paradigm in which participants first gave an initial intuitive response, under time pressure and cognitive load, and then gave a final response after deliberation. Results showed a clear training effect, as early as the initial (intuitive) stage. Immediately after training, most participants solved the problems correctly, without the need for a deliberation process. The findings confirm that the intuitive debiasing training effect extends to native French speakers., Competing Interests: The authors have no competing interests to declare., (Copyright: © 2024 The Author(s).)

Recent debiasing studies have shown that a short explanation about the correct solution to a reasoning problem can often improve performance of initially biased reasoners. Yet, with only one single training session, there is still a non-neglectable group of reasoners who remained biased. We explored whether repeated training on a battery of three reasoning tasks (i.e., bat-and-ball, base-rate neglect, and conjunction fallacy) can further boost reasoning performance. We recruited 120 adults, native English speakers, through Prolific Academic. We ran two studies with a battery of three classic reasoning tasks (see above). We used a two-response paradigm in which participants first gave an initial intuitive response, under time pressure and cognitive load, and then gave a final response after deliberation. In Study 1, we ran two repeated training sessions within one week. In Study 2, we ran a third training session two months after the initial study. Study 1 showed that after the first training session, most of the participants solved the problems correctly, as early as the initial intuitive stage. This training effect was further boosted by additional training, which helped almost the full sample to benefit. Study 2 indicated that these effects were robust and persisted after two months. The repetition of the training can further boost performance compared to the effect of one single training. These results are consistent with the wider literature on repeated testing and can serve as a proof-of-principle for a repeated debias training approach. • Debiasing training improves correct intuitive and deliberate reasoning performance. • Simply repeating the training is a straightforward tool to boost its efficacity. • The repetition also boosted metacognitive monitoring (i.e., error detection) processes. • The debias training effect is robust and persisted after two months. [ABSTRACT FROM AUTHOR]

• Exploit source knowledge to self-classify a target via self-supervised adaptation. • Implement double-time EMA to enable self-supervised target debiasing. • Extend morphological data to practical use with RiceDA dataset. • Conduct benchmarks and scenario deployment for real-world rice transportation. Industrial rice agriculture has transitioned towards intelligent computing, with an increasing likelihood of incorporating unmanned decision-making systems. However, one lingering issue that persists is the continued reliance on annotations for activity description. Applying transfer learning and domain-adaptive learning to non-labeled target data with the assistance of source data can be explicitly, but depends on mitigating domain discrepancy instead of principally generating predictive. This compromise both target robustness and comprehensibility. Our aim to enable a model to rely effectively on explicit source knowledge and self-labeling to reliably self-annotate. Our proposed self-supervised domain adaptation method provides debiasing by utilizing both former and present states, which reduces inconsistency when generating imbalanced pseudo-labels and implicitly utilizes source knowledge to consistently increase target-specific knowledge. Comprehensive results show improved performance by 7.5% for public data benchmarks and a gain of 2.4% for a rice transportation scenario. Ablation studies reveal improvements of up to 8.2% in Top-1 accuracy and 2.5% in Top-5 accuracy compared to an existing debiasing method, and highlights substantial parameter and FLOPs saving of up to 4.2 and 14 times, respectively. This underscores the feasibility of our system over self-label-dependent learning. [ABSTRACT FROM AUTHOR]

A major hurdle for the real time deployment of the AI models is ensuring trustworthiness of these models for the unseen population. More often than not, these complex models are black boxes in which promising results are generated. However, when scrutinized, these models begin to reveal implicit biases during the decision making, particularly for the minority subgroups. We develop an efficient adversarial de-biasing approach with partial learning by incorporating the existing concept activation vectors (CAV) methodology, to reduce racial disparities while preserving the performance of the targeted task. CAV is originally a model interpretability technique which we adopted to identify convolution layers responsible for learning race and only fine-tune up to that layer instead of fine-tuning the complete network, limiting the drop in performance The methodology has been evaluated on two independent medical image case-studies - chest X-ray and mammograms, and we also performed external validation on a different racial population. On the external datasets for the chest X-ray use-case, debiased models (averaged AUC 0.87) outperformed the baseline convolution models (averaged AUC 0.57) as well as the models trained with the popular fine-tuning strategy (averaged AUC 0.81). Moreover, the mammogram models is debiased using a single dataset (white, black and Asian) and improved the performance on an external datasets (averaged AUC 0.8 to 0.86) with completely different population (primarily Hispanic patients). In this study, we demonstrated that the adversarial models trained only with internal data performed equally or often outperformed the standard fine-tuning strategy with data from an external setting. The adversarial training approach described can be applied regardless of predictor's model architecture, as long as the convolution model is trained using a gradient-based method. We release the training code with academic open-source license - https://github.com/ramon349/JBI2023_TCAV_debiasing. [Display omitted] [ABSTRACT FROM AUTHOR]