國立虎尾科技大學 |

Essays on Responsible AI and Socially Equitable Computing.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Essays on Responsible AI and Socially Equitable Computing./
作者:	Boyarskaya, Margarita Shogen.
面頁冊數:	1 online resource (175 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-02, Section: B.
Contained By:	Dissertations Abstracts International86-02B.
標題:	Ethics. -
電子資源:	click for full text (PQDT)
ISBN:	9798383579794

Essays on Responsible AI and Socially Equitable Computing.
Boyarskaya, Margarita Shogen.

Essays on Responsible AI and Socially Equitable Computing. - 1 online resource (175 pages)

Source: Dissertations Abstracts International, Volume: 86-02, Section: B.

Thesis (Ph.D.)--New York University, 2024.

Includes bibliographical references

The focus of the present thesis manuscript is on socially equitable AI systems. The two aspects of this overarching theme that are of particular emphasis in the present essays are fair (anti-discriminatory) decision-making and theoretical assurance of data quality.AI fairness is a field that studies disparate or discriminatory effects of statistical prediction systems on different demographic groups. In the last decade, the subject of AI-enabled (un)fairness took center stage on the research agenda in Machine Learning. Simultaneously, it is the object of progressively more intense scrutiny and effort from qualitative socio-technical researches, as well as legal scholars and policy-makers. I focus on ensuring data quality, as well as on detection, diagnostics, and mitigation of bias in statistical decision-making. In addition to answering quantitative questions and examining methods for fair prediction, the essays present herein include elements of qualitative inquiry into larger questions about social implications of the use of AI. I use methods from social sciences both computational and qualitative to take on the hard questions in fairness (for example, the problem of disambiguating 'grey area' cases in which there is debate on whether using a proxy is impermissible).One of the leitmotifs of my work is the emphasis on the early and late stages of the AI-enabled decision pipelines oft-ignored or overshadowed by efforts targeted towards model-level adjustments (e.g., introducing a fairness constraint to optimization). In particular, I advocate and propose methods for high quality data collection, fair sampling, and equitable decision-making that occurs either before or after the prediction stage. Three examples of my work on these topics are the following:• My research in Essay 1 provides analytical expressions for the expected quality and cost of collecting data labels on crowdsourcing platforms using a strong version of majority voting. I put forth a full theoretical characterization of δ-margin voting a common method of aggregating votes into consensus labels. I propose an absorbing Markov chain model for the voting process and present closed-form formulas for the expected quality of the resulting data labels, as well as the probability density function, first, and second order moments for the workforce size required for reaching consensus. Further, the results enable an analytical comparison of different workforce groups: I show a way to adjust the consensus threshold so as to attain equal quality of results for processes assigned to lower vs. higher accuracy labor forces. In addition to enabling tighter control on both the quality of the training data and the costs associated with its collection, the findings of this essay have social implications for fair remuneration. The derived equation for the quality-equivalent payment ratio offers a possible solution to the problem of establishing relative wages depending on average response accuracy in each worker group. In other words, we provide a rule for determining how much a requester should pay for a task assigned to a group of higher-skilled workers relative to the wage paid for the same task to workers in a lower-skilled group, given that she constraints the resulting labels quality to be the same for both groups.• Essay 2 surveys the literature to introduce the reader to the field of AI Fairness. It collects, systematizes, and contrasts various definitions of fairness, highlighting theoretical and conceptual challenges implicit in each, as well as the impossibilities and challenges in establishing a broader comprehensive agenda that would inform fair AI deployment and governance. In the supplemental materials appended to this Essay, I further supply an original proposed taxonomy of various AI harms beyond unfairness.• The investigation presented in Essay 3 introduced missing clarity to the long-standing problem of unfairness via proxy variables. The Essay specifies the statistical relationships indirectly implied by three different conceptions of proxy variables native to law, economics, and CS fields, highlighting their incompatibility. I propose a concrete definition of the 'proxy variable problem' for the goals of algorithmic fairness. I survey existing statistical methods for addressing proxy variable use; map these methods onto the set of possible normative goals; and list legal tensions arising in each case. The Essay engages Causal Graphs methodology to identify how various stakeholders would incorporate different assumptions and objectives that give rise to (incompatible) modeling choices and thus different treatments of proxy variables. This helps explain why some proxy variables are commonly judged as being permissible to use, while other ones are almost unanimously agreed upon as being impermissible. The Essay puts forth a framework that could reconcile tensions and provide common theoretical ground for differing treatments of proxy variables. This is particularly useful in AI modeling use cases which can be described by very similar or identical causal graphs, yet solicit different decision-maker attitudes towards using or omitting a proxy. Our framework utilizes Directed Acyclic Graphs (DAGs) to explain agents' seemingly inconsistent modeling choices.An emphasis of my research work is on real-life, high impact settings with direct consequences for civic livelihood and changing industry standards. Of critical importance to my efforts is the dialogue with legal scholar community and with policymakers. In my fairness work, I emphasize legal constraints and affordances pertaining to the implementation of proposed technological solutions. My studies aim to inform algorithm design, AI managerial practice, private and public digital transformation strategy, and AI governance policy.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798383579794Subjects--Topical Terms:

555769
Ethics.
Subjects--Index Terms:

AI fairnessIndex Terms--Genre/Form:

554714
Electronic books.

Essays on Responsible AI and Socially Equitable Computing.
LDR:07227ntm a22004097 4500 001 1149847
005 20241022110530.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798383579794
035 $a (MiAaPQ)AAI31145542
035 $a AAI31145542
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Boyarskaya, Margarita Shogen. $3 1476212
245 1 0 $a Essays on Responsible AI and Socially Equitable Computing.
264 0 $c 2024
300 $a 1 online resource (175 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-02, Section: B.
500 $a Advisor: Ipeirotis, Panos G.
502 $a Thesis (Ph.D.)--New York University, 2024.
504 $a Includes bibliographical references
520 $a The focus of the present thesis manuscript is on socially equitable AI systems. The two aspects of this overarching theme that are of particular emphasis in the present essays are fair (anti-discriminatory) decision-making and theoretical assurance of data quality.AI fairness is a field that studies disparate or discriminatory effects of statistical prediction systems on different demographic groups. In the last decade, the subject of AI-enabled (un)fairness took center stage on the research agenda in Machine Learning. Simultaneously, it is the object of progressively more intense scrutiny and effort from qualitative socio-technical researches, as well as legal scholars and policy-makers. I focus on ensuring data quality, as well as on detection, diagnostics, and mitigation of bias in statistical decision-making. In addition to answering quantitative questions and examining methods for fair prediction, the essays present herein include elements of qualitative inquiry into larger questions about social implications of the use of AI. I use methods from social sciences both computational and qualitative to take on the hard questions in fairness (for example, the problem of disambiguating 'grey area' cases in which there is debate on whether using a proxy is impermissible).One of the leitmotifs of my work is the emphasis on the early and late stages of the AI-enabled decision pipelines oft-ignored or overshadowed by efforts targeted towards model-level adjustments (e.g., introducing a fairness constraint to optimization). In particular, I advocate and propose methods for high quality data collection, fair sampling, and equitable decision-making that occurs either before or after the prediction stage. Three examples of my work on these topics are the following:• My research in Essay 1 provides analytical expressions for the expected quality and cost of collecting data labels on crowdsourcing platforms using a strong version of majority voting. I put forth a full theoretical characterization of δ-margin voting a common method of aggregating votes into consensus labels. I propose an absorbing Markov chain model for the voting process and present closed-form formulas for the expected quality of the resulting data labels, as well as the probability density function, first, and second order moments for the workforce size required for reaching consensus. Further, the results enable an analytical comparison of different workforce groups: I show a way to adjust the consensus threshold so as to attain equal quality of results for processes assigned to lower vs. higher accuracy labor forces. In addition to enabling tighter control on both the quality of the training data and the costs associated with its collection, the findings of this essay have social implications for fair remuneration. The derived equation for the quality-equivalent payment ratio offers a possible solution to the problem of establishing relative wages depending on average response accuracy in each worker group. In other words, we provide a rule for determining how much a requester should pay for a task assigned to a group of higher-skilled workers relative to the wage paid for the same task to workers in a lower-skilled group, given that she constraints the resulting labels quality to be the same for both groups.• Essay 2 surveys the literature to introduce the reader to the field of AI Fairness. It collects, systematizes, and contrasts various definitions of fairness, highlighting theoretical and conceptual challenges implicit in each, as well as the impossibilities and challenges in establishing a broader comprehensive agenda that would inform fair AI deployment and governance. In the supplemental materials appended to this Essay, I further supply an original proposed taxonomy of various AI harms beyond unfairness.• The investigation presented in Essay 3 introduced missing clarity to the long-standing problem of unfairness via proxy variables. The Essay specifies the statistical relationships indirectly implied by three different conceptions of proxy variables native to law, economics, and CS fields, highlighting their incompatibility. I propose a concrete definition of the 'proxy variable problem' for the goals of algorithmic fairness. I survey existing statistical methods for addressing proxy variable use; map these methods onto the set of possible normative goals; and list legal tensions arising in each case. The Essay engages Causal Graphs methodology to identify how various stakeholders would incorporate different assumptions and objectives that give rise to (incompatible) modeling choices and thus different treatments of proxy variables. This helps explain why some proxy variables are commonly judged as being permissible to use, while other ones are almost unanimously agreed upon as being impermissible. The Essay puts forth a framework that could reconcile tensions and provide common theoretical ground for differing treatments of proxy variables. This is particularly useful in AI modeling use cases which can be described by very similar or identical causal graphs, yet solicit different decision-maker attitudes towards using or omitting a proxy. Our framework utilizes Directed Acyclic Graphs (DAGs) to explain agents' seemingly inconsistent modeling choices.An emphasis of my research work is on real-life, high impact settings with direct consequences for civic livelihood and changing industry standards. Of critical importance to my efforts is the dialogue with legal scholar community and with policymakers. In my fairness work, I emphasize legal constraints and affordances pertaining to the implementation of proposed technological solutions. My studies aim to inform algorithm design, AI managerial practice, private and public digital transformation strategy, and AI governance policy.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Ethics. $3 555769
650 4 $a Computer science. $3 573171
653 $a AI fairness
653 $a Data quality
653 $a Majority voting
653 $a Responsible AI
653 $a Vote aggregation
655 7 $a Electronic books. $2 local $3 554714
690 $a 0984
690 $a 0394
690 $a 0796
690 $a 0800
710 2 $a New York University. $b Information Systems. $3 1473601
710 2 $a ProQuest Information and Learning Co. $3 1178819
773 0 $t Dissertations Abstracts International $g 86-02B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31145542 $z click for full text (PQDT)