Дисертації з теми "Mitigation des biais"

Il est de notoriété que les modèles d'apprentissage profond sont sensibles aux biais qui peuvent être présents dans les données utilisées pour l'apprentissage. Ces biais qui peuvent être définis comme de l'information inutile ou préjudiciable pour la tâche considérée, peuvent être de différentes natures: on peut par exemple trouver des biais dans les styles d'écriture utilisés, mais aussi des biais bien plus problématiques portant sur le sexe ou l'origine ethnique des individus. Ces biais peuvent provenir de différentes sources, comme des annotateurs ayant créé les bases de données, ou bien du processus d'annotation lui-même. Ma thèse a pour sujet l'étude de ces biais et, en particulier, s'organise autour de la mitigation des effets des biais sur l'apprentissage des modèles de Traitement Automatique des Langues (TAL). J'ai notamment beaucoup travaillé avec les modèles pré-entraînés comme BERT, RoBERTa ou UnifiedQA qui sont devenus incontournables ces dernières années dans tous les domaines du TAL et qui, malgré leur large pré-entraînement, sont très sensibles à ces problèmes de biais. Ma thèse s'organise en trois volets, chacun présentant une façon différente de gérer les biais présents dans les données. Le premier volet présente une méthode permettant d'utiliser les biais présents dans une base de données de résumé automatique afin d'augmenter la variabilité et la contrôlabilité des résumés générés. Puis, dans le deuxième volet, je m'intéresse à la génération automatique d'une base de données d'entraînement pour la tâche de question-réponse à choix multiples. L'intérêt d'une telle méthode de génération est qu'elle permet de ne pas faire appel à des annotateurs et donc d'éliminer les biais venant de ceux-ci dans les données. Finalement, je m'intéresse à l'entraînement d'un modèle multitâche pour la reconnaissance optique de texte. Je montre dans ce dernier volet qu'il est possible d'augmenter les performances de nos modèles en utilisant différents types de données (manuscrites et tapuscrites) lors de leur entraînement
It is well known that deep learning models are sensitive to biases that may be present in the data used for training. These biases, which can be defined as useless or detrimental information for the task in question, can be of different kinds: one can, for example, find biases in the writing styles used, but also much more problematic biases relating to the sex or ethnic origin of individuals. These biases can come from different sources, such as annotators who created the databases, or from the annotation process itself. My thesis deals with the study of these biases and, in particular, is organized around the mitigation of the effects of biases on the training of Natural Language Processing (NLP) models. In particular, I have worked a lot with pre-trained models such as BERT, RoBERTa or UnifiedQA which have become essential in recent years in all areas of NLP and which, despite their extensive pre-training, are very sensitive to these bias problems.My thesis is organized in three parts, each presenting a different way of managing the biases present in the data. The first part presents a method allowing to use the biases present in an automatic summary database in order to increase the variability and the controllability of the generated summaries. Then, in the second part, I am interested in the automatic generation of a training dataset for the multiple-choice question-answering task. The advantage of such a generation method is that it makes it possible not to call on annotators and therefore to eliminate the biases coming from them in the data. Finally, I am interested in training a multitasking model for optical text recognition. I show in this last part that it is possible to increase the performance of our models by using different types of data (handwritten and typed) during their training

Automated decision support tools are designed to aid users and improve their performance in certain tasks by providing advice in the form of prompts, alarms, assessments, or recommendations. However, recent evidence suggests that sometimes use of such tools introduces decision errors that are not made without the tool. We refer to this phenomenon as “automation bias” (AB), resulting in a broader definition of this term than used by many authors. Sometimes, such automation-induced errors can even result in overall performance (in terms of correct decisions) which is actually worse with the tool than without it. Our literature review reveals an emphasis on mediators affecting automation bias and some mitigation strategies aimed at reducing it. However, there is a lack of research on the cognitive causal explanations for automation bias and on adaptive mitigation strategies that result in tools that adapt to the needs and characteristics of individual users. This thesis aims to address some of these gaps in the literature and focuses on systems consisting of a human and an automated tool which does not replace, but instead supports the human towards making a decision, with the overall responsibility lying with the human user. The overall goal of this thesis is to help reduce the rate of automation bias through a better understanding of its causes and the proposal of innovative, adaptive mitigation strategies. To achieve this, we begin with an extensive literature review on automation bias including examples, mediators, explanations, and mitigations while identifying areas for further research. This review is followed by the presentation of three experiments aimed at reducing the rate of AB in different ways: (1) an experiment to explore causal mechanisms of automation bias, the effect of the mere presence of tool advice before its presentation and the effect of the sequence of tool advice in a glaucoma risk calculator environment, (2) simulations that apply concepts of diversity to human + human systems to improve system performance in a breast cancer double reading programme, and (3) an experiment to study the possibility of improving system performance by tailoring tool setting (sensitivity / specificity combination) for groups of similarly skilled users and cases of similar difficulty level using a spellchecking tool. Results from the glaucoma experiment provide evidence of the effect of the presence of tool advice on user decisions - even before its presentation, as well as evidence of a newly introduced cognitive mechanism (users’ strategic change in decision threshold) which may account for some automation bias errors previously observed but unexplained in the literature. Results from the double reading experiment provide evidence of the benefits of diversity in improving system performance. Finally, results from the spell checker experiment provide evidence that groups of similarly skilled users perform better at different tool settings, that the same group of users perform better using a different tool setting in difficult versus easy tasks, and that use of simple models of user behaviour may allow the prediction, among a subset of tool settings for a certain tool, the tool setting that would be most appropriate for each user ability group and class of case difficulty.

Artificial Intelligence (AI) systems are increasingly used in society to make decisions that can have direct implications on human lives; credit risk assessments, employment decisions and criminal suspects predictions. As public attention has been drawn towards examples of discriminating and biased AI systems, concerns have been raised about the fairness of these systems. Face recognition systems, in particular, are often trained on non-diverse data sets where certain groups often are underrepresented in the data. The focus of this thesis is to provide insights regarding different aspects that are important to consider in order to mitigate algorithmic bias as well as to investigate the practical implications of bias in AI systems. To fulfil this objective, qualitative interviews with academics and practitioners with different roles in the field of AI and a quantitative online survey is conducted. A practical scenario covering face recognition and gender bias is also applied in order to understand how people reason about this issue in a practical context. The main conclusion of the study is that despite high levels of awareness and understanding about challenges and technical solutions, the academics and practitioners showed little or no awareness of legal aspects regarding bias in AI systems. The implication of this finding is that AI can be seen as a disruptive technology, where organizations tend to develop their own mitigation tools and frameworks as well as use their own moral judgement and understanding of the area instead of turning to legal authorities.

Magnetic interference in the motion capture environment is caused primarily by ferromagnetic objects and current-carrying devices disturbing the ambient, geomagnetic field. Inertial sensors gather magnetic data to determine and stabilize their global heading estimates, and such magnetic field disturbances alter heading estimates. This decreases orientation accuracy and therefore decreases motion capture accuracy. The often used Kalman Filter approach deals with magnetic interference by ignoring the magnetic data during periods interference is encountered, but this method is only effective when the disturbances are ephemeral, and cannot not retroactively repair data from disturbed time periods. The objective of this research is to develop a method of magnetic interference mitigation for environments where magnetic interference is the norm rather than the exception. To the knowledge of this author, the ability to use inertial and magnetic sensors to capture accurate, global, and drift-free orientation data in magnetically disturbed areas has yet to be developed. Furthermore there are no methods known to this author that are able to use data from undisturbed time periods to retroactively repair data from disturbed time periods. The investigation begins by exploring the use of magnetic shielding, with the reasoning that application of shielding so as to impede disturbed fields from affecting the inertial sensors would increase orientation accuracy. It was concluded that while shielding can mitigate the effect of magnetic interference, its application requires a tedious trial and error testing that was not guaranteed to improve results. Furthermore, shielding works by redirecting magnetic field lines, increasing field complexity, and thus has a high potential to exacerbate magnetic interference. Shielding was determined to be an impractical approach, and development of a magnetic inference mitigation algorithm began. The algorithm was constructed such that magnetic data would be filtered before inclusion in the orientation estimate, with the result that exposure in an undisturbed environment would improve estimation, but exposure to a disturbed environment would have no effect. The algorithm was designed for post-processing, rather than real-time use as Kalman Filters are, which enabled magnetic data gathered before and after a time point could affect estimation. The algorithm was evaluated by comparing it with the Kalman Filter approach of the company XSENS, using the gold standard of optical motion capture as the reference point. Under the tested conditions of stationary periods and smooth planar motion, the developed algorithm was resistant to magnetic interference for the duration of testing, while the Kalman Filter began to degrade after approximately 15 seconds. In a 190 second test, of which 180 were spent in a disturbed environment, the developed algorithm resulted in 0.4 degrees of absolute error, compared to the of the Kalman Filter’s 78.8 degrees. The developed algorithm shows the potential for inertial systems to be used effectively in situations of consistent magnetic interference. As the benefits of inertial motion capture make it a more attractive option than optical motion capture, immunity to magnetic interference significantly expands the usable range of motion capture environments. Such expansion would be beneficial for motion capture studies as a whole, allowing for the cheaper, more practical inertial approach to motion capture to supplant the more expensive and time consuming optimal option.

All computer simulation models require some form of initialization before their outputs can be considered meaningful. Simulation models are typically initialized in a particular, often "empty" state and therefore must be "warmed-up" for an unknown amount of simulation time before reaching a "quasi-steady-state" representative of the systems' performance. The portion of the output series that is influenced by the arbitrary initialization is referred to as the initial transient and is a widely recognized problem in simulation analysis. Although several methods exist for removing the initial transient, there are no methods that perform well in all applications. This research evaluates the effectiveness of several techniques for reducing initialization bias from simulations using the commercial transportation simulation model VISSIM®. The three methods ultimately selected for evaluation are Welch's Method, the Marginal Standard Error Rule (MSER) and the Volume Balancing Method currently being used by the CORSIM model. Three model instances - a single intersection, a corridor, and a large network - were created to analyze the length of the initial transient for varying scenarios, under high and low demand scenarios. After presenting the results of each initialization method, advantages and criticisms of each are discussed as well as issues that arose during the implementation. The results for estimation of the extent of the initial transient are compared across each method and across the varying model sizes and volume levels. Based on the results of this study, Welch's Method is recommended based on is consistency and ease of implementation.

This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Thesis: M. Eng., Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science, 2019
Cataloged from student-submitted PDF version of thesis.
Includes bibliographical references (pages 81-84).
Machine Learning is becoming more and more influential in our society. Algorithms that learn from data are streamlining tasks in domains like employment, banking, education, heath care, social media, etc. Unfortunately, machine learning models are very susceptible to unintended bias, resulting in unfair and discriminatory algorithms with the power to adversely impact society. This unintended bias is usually subtle, emanating from many different sources and taking on many forms. This thesis will focus on understanding how unfair biases with respect to various demographic groups show up in machine learning systems. Furthermore, we develop multiple techniques to mitigate unintended demographic bias at various stages of typical machine learning pipelines. Using Natural Language Processing as a framework, we show substantial improvements in fairness for standard machine learning systems, when using our bias mitigation techniques.
by Christopher Sweeney.
M. Eng.
M.Eng. Massachusetts Institute of Technology, Department of Electrical Engineering and Computer Science

Inverters have traditionally been used mostly in standalone systems (non-grid connected), Uninterruptible Power Supplies (UPS) and, more recently, in distributed generated systems (DGs). DG systems are based on grid connected inverters and are increasingly being connected to utility grids to convert renewable energy sources to the utility grids. Such sources are likely to have a significant impact in the future in meeting the electricity demands of industry and domestic consumption. Common DGs utilize DC power sources such as fuel cells, batteries, photovoltaic (solar) power, and wind power. Most of power supplies in domestic and industries are AC power consumers and, for this reason, the DC power has to be converted to meet the requirement. Two main causes of DC current in inverter power transformer are: 1) Non-linearity and asymmetry in its switching mechanism which may result in undesired DC current at its input. This DC current introduced into an inverter transformer results in the transformer's magnetic flux distortion and in some cases magnetic saturation. This, in turn, results in asymmetrical primary currents in the transformer (inverter side). This is due to the non linear characteristics of the transformer magnetic flux. 2) The same effects can be produced by the connection of asymmetrical loads (e.g. asymmetrical rectifier) to the inverter output. The result in both cases is an asymmetrical magnetic flux in the transformer. This is manifested as even and odd current harmonics as well as an increase in the reactive power requirement from the inverter. vi To remedy this situation, it is, therefore, necessary to incorporate into the inverter's control system a mechanism of cancelling the DC magnetic motive force (mmf) that causes the magnetic flux distortion. This Thesis presents a method of introducing a DC voltage component in the inverter's voltage output so as to inject the necessary DC current into the primary side of the inverter's transformer so as to cancel the total DC mmf that the transformer is subjected to ( supply and load side). This project consists of three main parts namely: Modeling, Simulation and Laboratory Experiment. Activities undertaken under Modeling and Simulation were as follows: Determining the effects of DC current on a power transformer. Investigating the likely occurrence of saturation of the power transformer incorporated in inverter systems. Mitigating the effects that can be caused by the presence of a DC component in the windings of a power transformer. After understanding the literature on the subject of interest, MATLAB SIMULINK and MATLAB m-files were used to simulate the behavior of the power transformer under three situations : The transformer under linear load. The transformer subjected to asymmetrical loading. The inverter system that has a power transformer on its output were designed in MATLAB and used to simulate the situation for each case. To validate the theory and simulation results, experimental work was carried out as follows: vii Investigation of the effects that DC (current) injection can have on a 6 kVA power transformer. Investigation of the performance of a 6 kVA power transformer under linear loading. Investigation of the performance of a 6 kVA power transformer under non-linear loads. Investigation of the likely occurrence of DC offset in inverter system. Mitigation of the effect of DC bias on power transformer using extra windings. Mitigation of the effects of DC offset in power inverter transformer by using the second harmonic content of the primary current as a feedback signal. Results obtained showed a successful implementation of the proposed method. However limitations of the controller performances were experienced and will require future work. It was concluded that a total removal of the undesired effects of DC bias is achievable and that total removal of DC offset in power inverter transformer is possible if the limitations of the controller are overcome.

Fingerprint identification is a discipline used within forensic science which assists in criminal investigations1, 2. The process of fingerprint identification involves the comparison of crime scene evidence with known exemplars. This form of examination is heavily reliant on human examiners and their conclusions as to whether there is an identification, exclusion or insufficient information to identify3. This form of forensic identification has become a focus due to concern of the effects of cognitive bias on examiners conclusions. Concerns have prompted research into the area of approaches to mitigate bias throughout forensic fingerprint protocols. Research into the common sources of bias during a fingerprint examination was conducted to gain an understanding of how bias may potentially be reduced. Throughout this dissertation the psychological and forensic approaches to bias were reviewed and the international and Australian approaches to bias mitigation were discussed. This found that there was evidence of a widespread issue regarding human cognitive bias in fingerprint examiners, however, there were no uniform mitigation strategies in place. Limitations to recommended approaches and currently implemented strategies have been reviewed, identifying that there is still a need for further research into the theoretical approaches to overcome bias. Therefore, leading to the formation of a study that aims to identify the theoretical approaches as suggested by literature, and critically review the effectiveness of these methods in controlling and reducing bias. The potential outcome from the suggested study may result in a useful document that will provide the practical field of forensic science with a comprehensive and critical review of approaches to assist in the development of standardised protocols.

Gender differences in salaries are prominent in most fields. Several laws exist to decrease the amount of pay discrimination and provide remedies for discriminatory organizational behaviors, yet these laws have proven insufficient to eradicate pay inequities. One source for such discrimination in pay stems from the evaluation of employee performance. Performance appraisal systems can be biased in very small ways that yield larger negative effects on later employment-related decisions, such as compensation. The goal of this study was to examine decision-making processes and conclusions raters make during the evaluation of employees. It was expected that the type of presentation and the content of the ratings of performance sub-dimensions would affect gender differences in composite ratings, salary increases, and merit bonuses. Specifically, women were expected to be rated lower when employee performance information was presented sequentially, where it would be harder to directly compare one employee with another and thus not mitigate sex bias. Comparatively, when employee performance information was presented in aggregate form, where comparisons among employees would be easier, no sex bias was expected. It was also hypothesized that in the sequential condition, participants would provide casuistry-based reasoning for their decisions such that explanations for men's better performance would be justified by their performance on the sub-dimension on which the male candidate was rated highly. No effect was found for target gender on any of the outcomes. There was a significant difference for participant gender in the amount of salary increases and merit bonuses assigned. Male participants assigned significantly higher raises and bonuses than female participants to employees. There was also a strong main effect for task-related skills across all outcomes. Employees who were higher in the task dimension were rated higher, awarded higher pay, and given larger bonuses. There were no gender differences in the task ratings. Qualitative data analyses support these conclusions. Although participants provided lengthy reasons for their decisions, none showed or explained a gender difference. Limitations and recommendations for future studies are discussed.

This work has been undertaken within the frame of the FPI scholarship BES-2009-028505 of 30/07/2009, by the "Secretario de Estado de Investigación del Ministerio de Ciencia e Innovación", related to the project TEC2008-06764-C02-01 (Universitat Politècnica de Catalunya) titled "Advanced concepts on active and passive microwave remote sensing: technology and applications". In a more general scope, this thesis is related to the Remote Sensing Laboratory (Signal Theory & Communication Department, UPC) on-going activities, within the SMOS (Soil Moisture and Ocean Salinity) mission by the European Space Agency (ESA). These activities have been organized to provide original advances in the following four main topics: 1) SMOS system performance assessment. SMOS commissioning phase finalized in May 2010 providing preliminary performance results. Therefore, as a first step in this work, radiometric sensitivity and spatial bias (systematic spatial errors) have been reviewed and assessed. To achieve this, new techniques have been developed to better compare SMOS images to ground truth targets. SMOS ocean views have been selected as the best option since provide smooth brightness temperature distributions for which accurate models are available. 2) Spatial bias mitigation techniques. Spatial bias was a major issue in SMOS first images since it was larger than expected. In order to identify the dominant sources of spatial bias, each potential source of error, at calibration level,has been simulatedand compared tomeasured SMOS error patterns. 3) Spatial bias in the EAF-FoV. Whereas SMOS performance in the alias free FoV was well within mission requirements, this issue was especially important for the four polarimetric temperatures in the extended regions. This issuehas been addressed in two ways. First a new approach has been developed to systematically analyze spatial bias by splitting SMOS EAF-FoV into different regions of interest. Also, the pixels have been arranged per angle of incidence. Secondly, the impact in each region of calibration errors and different image reconstruction has been addressed. 4) Spatial bias in full-pol measurement.The main objective of this work has been achieved by developing, implementing and validating a SMOS full-pol G-matrix image reconstruction procedure that resulted into significantly improved polarimetric images. This tool allowed a comprehensive assessment of the two error contributions that have been identified as the main sources of residual spatial bias: the so-called floor error and antenna pattern uncertainty. Finally, as a validation of the quality of SMOS Level 1 fully polarimetric data that has been achieved, the last chapter of this work has been devoted to provide operational Faraday rotation retrievals in a per snap-shot basis.

Thesis (M.S. in Information Systems and Operations)--Naval Postgraduate School, March 2003.
Thesis advisor(s): Steven J. Iatrou, Karen Guttieri. Includes bibliographical references (p. 67-68). Also available online.

abstract: Exabytes of data are created online every day. This deluge of data is no more apparent than it is on social media. Naturally, finding ways to leverage this unprecedented source of human information is an active area of research. Social media platforms have become laboratories for conducting experiments about people at scales thought unimaginable only a few years ago. Researchers and practitioners use social media to extract actionable patterns such as where aid should be distributed in a crisis. However, the validity of these patterns relies on having a representative dataset. As this dissertation shows, the data collected from social media is seldom representative of the activity of the site itself, and less so of human activity. This means that the results of many studies are limited by the quality of data they collect. The finding that social media data is biased inspires the main challenge addressed by this thesis. I introduce three sets of methodologies to correct for bias. First, I design methods to deal with data collection bias. I offer a methodology which can find bias within a social media dataset. This methodology works by comparing the collected data with other sources to find bias in a stream. The dissertation also outlines a data collection strategy which minimizes the amount of bias that will appear in a given dataset. It introduces a crawling strategy which mitigates the amount of bias in the resulting dataset. Second, I introduce a methodology to identify bots and shills within a social media dataset. This directly addresses the concern that the users of a social media site are not representative. Applying these methodologies allows the population under study on a social media site to better match that of the real world. Finally, the dissertation discusses perceptual biases, explains how they affect analysis, and introduces computational approaches to mitigate them. The results of the dissertation allow for the discovery and removal of different levels of bias within a social media dataset. This has important implications for social media mining, namely that the behavioral patterns and insights extracted from social media will be more representative of the populations under study.
Dissertation/Thesis
Doctoral Dissertation Computer Science 2017

碩士
國立臺北大學
自然資源與環境管理研究所
107
When applying the contingent value method (CVM) to value non-marketed goods (such as wetlands), researchers often face the problem of hypothetical bias (HB) in that respondents’ answered willingness to pay (WTP) is often different from their real WTP. The methods to eliminate HB are divided into ex post calibration and ex ante reminder. The former compare the real market data and questionnaire data to adjust the WTP after the survey undertaken; the latter is to provide some information in questionnaire before inquiring respondents, so that respondents can correct their WTP before answering. CT is a common practice in the ex-ante reminder. In the context of CT, the existence of HB is explicitly explained, respondents are reminded that they may fill in a dollar amount that does not match their real WTP, and then respondents are asked to promise to fill in a WTP that is closer to their real one. In order to investigate people’s WTP for conserving the Qijiawanxi wetland, Shuanglian wetland and wetlands nationwide, this study interviewed 600 questionnaires in person and added 300 short-length CTs to test whether CT could reduce HB. Three models (OLS model, Tobit model, and sample selection model) are used to analyze the empirical WTP bid functions, the results show that the effect of CT on reducing HB is not statistically significant. The reasons for this result may include: (1) The CT taken by this study is a short narrative, which may not be enough for the respondents to understand what is called HB; (2) Establishing an environment mental accounting framework (MAF) for the respondents in questionnaire can reduce the HB. First, several instruction cards the clearly state the overall scope of the wetlands surveyed are presented and explained to respondents. Secondly, we inquiry respondents’ WTP for preserving all wetlands in Taiwan, then asks their WTP for individual wetlands. (3) HB and other biases such as hypothetical and embedding biased are overlapped, the MAF for reducing the embedding bias also help to reduce HB.

碩士
國立臺灣大學
圖書資訊學研究所
105
Confirmation bias, the preferential seeking of confirmatory information, can become an obstacle for the disseminate of valid health information online and lead to biased decision. Following previous findings, preference-inconsistent recommendation can be used to overcome this bias. We conducted an experiment to study the impact of expert system and popularity system on mitigating confirmation bias, the confirmation bias was measured in the selection phase, the evaluation phase and the final decision phases. 78 participants aged 40-70 were recruited. Participants were informed that they would participate in a health information experiment involving two cancer screening debates. Participants were assigned in such a way that each participant would see two fairly different interfaces for the two tasks. We found that the evaluation bias and the final decision were more persistent than the selection bias. The comparison between the two systems revealed that expert system has a better mitigation effect than the popularity system in the selection bias. Furthermore, it was observed that expert system would have better mitigation effect on high-involvement issue. We also found strong gender difference in our experiment. Future study which aims to investigate the mitigation effect of different techniques should take gender as a confounding variable and choose health issues which are more comparable.

Biometric verification refers to the automatic verification of a person’s identity based on their behavioural and biological characteristics. Among various biometric modalities, the face is one of the most widely used since it is easily acquirable in unconstrained environments and provides a strong uniqueness. In recent years, face recognition systems spread world-wide and are increasingly involved in critical decision-making processes such as finance, public security, and forensics. The growing effect of these systems on everybody’s daily life is driven by the strong enhancements in their recognition performance. The advances in extracting deeply-learned feature representations from face images enabled the high-performance of current face recognition systems. However, the success of these representations came at the cost of two major discriminatory concerns. These concerns are driven by soft-biometric attributes such as demographics, accessories, health conditions, or hairstyles. The first concern is about bias in face recognition. Current face recognition solutions are built on representation-learning strategies that optimize total recognition performance. These learning strategies often depend on the underlying distribution of soft-biometric attributes in the training data. Consequently, the behaviour of the learned face recognition solutions strongly varies depending on the individual’s soft-biometrics (e.g. based on the individual’s ethnicity). The second concern tackles the user’s privacy in such systems. Although face recognition systems are trained to recognize individuals based on face images, the deeply-learned representation of an individual contains more information than just the person’s identity. Privacy-sensitive information such as demographics, sexual orientation, or health status, is encoded in such representations. However, for many applications, the biometric data is expected to be used for recognition only and thus, raises major privacy issues. The unauthorized access of such individual’s privacy-sensitive information can lead to unfair or unequal treatment of this individual. Both issues are caused by the presence of soft-biometric attribute information in the face images. Previous research focused on investigating the influence of demographic attributes on both concerns. Consequently, the solutions from previous works focused on the mitigation of demographic-concerns only as well. Moreover, these approaches require computationally-heavy retraining of the deployed face recognition model and thus, are hardly-integrable into existing systems. Unlike previous works, this thesis proposes solutions to mitigating soft-biometric driven bias and privacy concerns in face recognition systems that are easily-integrable in existing systems and aim for more comprehensive mitigation, not limited to pre-defined demographic attributes. This aims at enhancing the reliability, trust, and dissemination of these systems. The first part of this work provides in-depth investigations on soft-biometric driven bias and privacy concerns in face recognition over a wide range of soft-biometric attributes. The findings of these investigations guided the development of the proposed solutions. The investigations showed that a high number of soft-biometric and privacy-sensitive attributes are encoded in face representations. Moreover, the presence of these soft-biometric attributes strongly influences the behaviour of face recognition systems. This demonstrates the strong need for more comprehensive privacy-enhancing and bias-mitigating technologies that are not limited to pre-defined (demographic) attributes. Guided by these findings, this work proposes solutions for mitigating bias in face recognition systems and solutions for the enhancement of soft-biometric privacy in these systems. The proposed bias-mitigating solutions operate on the comparison- and score-level of recognition system and thus, can be easily integrated. Incorporating the notation of individual fairness, that aims at treating similar individuals similarly, strongly mitigates bias of unknown origins and further improves the overall-recognition performance of the system. The proposed solutions for enhancing the soft-biometric privacy in face recognition systems either manipulate existing face representations directly or changes the representation type including the inference process for verification. The manipulation of existing face representations aims at directly suppressing the encoded privacy-risk information in an easily-integrable manner. Contrarily, the inference-level solutions indirectly suppress this privacy-risk information by changing the way of how this information is encoded. To summarise, this work investigates soft-biometric driven bias and privacy concerns in face recognition systems and proposed solutions to mitigate these. Unlike previous works, the proposed approaches are (a) highly effective in mitigating these concerns, (b) not limited to the mitigation of concerns origin from specific attributes, and (c) easily-integrable into existing systems. Moreover, the presented solutions are not limited to face biometrics and thus, aim at enhancing the reliability, trust, and dissemination of biometric systems in general.

A number of studies have indicated that Self-Affirmation (SA) manipulation plays a role in decreasing an attention bias toward threatening information, especially when there is a threat to the self. Despite this, to date no studies have investigated if SA can mitigate the attention bias toward socially threatening stimuli that is exhibited in individuals high in social anxiety. The aim of the current study is to explore the possible moderating effects of SA manipulation on the attention bias in individuals with high social anxiety. 150 participants completed the study, 8 of which were excluded based upon poor accuracy on the emotional Stroop task. This left a sample of data from 142 participants (aged 18-71 years, M=26.66, SD=l0.94). Participants were required to complete a number of measures online, including the Brief Fear of Negative Evaluation Scale (Leary,1983) and the Beck Depression Inventory- 2nd edition (Beck, Steer & Brown, 1996) as well as undergo a SA manipulation or control condition similar to that used in previous SA studies (Armitage, Harris and Arden, 2011). The independent variables were social anxiety group (high, low) and experimental condition (SA manipulation, control). The dependent variable was emotional interference score as identified by latency times to identify print colour for social threat and control words presented in an emotional Stroop task. A 2 by 2 (Social anxiety: high, low; experimental condition: SA, control; covariate: depression score) ANCOV A was run on the data and revealed no significant main effects for social anxiety group or experimental condition (p>.05). The social anxiety group x experimental condition interaction was also not significant (p>.05). Unfortunately, the aim of the study could not be tested due to no significant differences in emotional interference scores between those high and low in social anxiety. This raises questions about the presence of an attention bias toward socially threatening words by individuals high in social anxiety in an online environment.

Conceptual design is a series of complex cognitive processing tasks and research seeking to further understand design cognition will benefit by considering literature from the field of psychology. This thesis presents two research projects, which sought to understand and mitigate design biases in conceptual design through the application of theories from biological and cognitive psychology. The first of these puts forward a novel model of design creativity based on connectionist theory and a neurological phenomenon known as long-term potentiation. This model is applied to provide new insights into design fixation and develop interventions to assist designers overcome fixation. The second project seeks to establish that cognitive heuristics and biases predictably influence design cognition. Two studies are discussed that examined the role of confirmation bias in design. The first establishes that confirmation bias is present during concept generation; the second demonstrates that decision matrices can mitigate confirmation bias in concept evaluation.