2023-04-24

Timnit Gebru and Margaret Mitchell on AI Accountability

2023.4.24: news.cyb/ai/
Timnit Gebru and Margaret Mitchell on AI Accountability

. late 2020 and early 2021, two researchers 

— Timnit Gebru and Margaret Mitchell — 

were given special attention

after they authored a research paper

addressing flaws in today's AI systems.

[James Vincent Apr 19, 2023]

https://www.theverge.com/2023/4/19/23689554/google-ai-chatbot-bard-employees-criticism-pathological-liar

some papers by Timnit Gebru and Margaret Mitchell:


2022:

Vinodkumar Prabhakaran, 

Margaret Mitchell, 

Timnit Gebru,

Iason Gabriel 

https://arxiv.org/abs/2210.02667

A Human Rights-Based Approach to Responsible AI.


. Research on fairness, accountability, transparency and ethics

of AI-based interventions in society 

has gained much-needed momentum in recent years. 

However it lacks an explicit alignment with 

a set of normative values and principles 

that guide this research and interventions. 

Rather, an implicit consensus is often assumed to hold 

for the values we impart into our models

- something that is at odds with the pluralistic world we live in. 

In this paper, we put forth the doctrine of

universal human rights

as a set of globally salient and cross-culturally recognized

set of values that can serve as a grounding framework for explicit value alignment in responsible AI

- and discuss its efficacy as a framework for 

civil society partnership and participation. 

We argue that a human rights framework orients the research in this space 

away from the machines

and the risks of their biases, 

and towards humans 

and the risks to their rights, 

essentially helping to center the conversation

around who is harmed, what harms they face, 

and how those harms may be mitigated.


2021:

Daniel J. Liebling Katherine Heller 

Margaret Mitchell

Mark Díaz Michal Lahav 

Niloufar Salehi Samantha Robertson Samy Bengio

Timnit Gebru

Wesley Deng

https://research.google/pubs/pub50504/

Three Directions for the Design of

Human-Centered Machine Translation.


https://www.samantha-robertson.com/publication/hcmt/hcmt.pdf

As people all over the world adopt

machine translation (MT) to communicate across languages, 

there is increased need for affordances that 

aid users in understanding when to rely on

automated translations. 

Identifying the information and interactions that will

most help users meet their translation needs

is an open area of research at the intersection of

Human-Computer Interaction (HCI) 

and Natural Language Processing (NLP). 

This paper advances work in this area by 

drawing on a survey of users' strategies in

assessing translations. 

We identify three directions for the 

design of translation systems that support

more reliable and effective use of machine translation: 

helping users craft good inputs, 

helping users understand translations, 

and expanding interactivity and adaptivity. 

We describe how these can be introduced in current MT systems

and highlight open questions for HCI and NLP research.


2020:

Andrew Smart Becky White Ben Hutchinson Daniel Theron Inioluwa Deborah Raji Jamila Smith-Loud

Margaret Mitchell

Parker Barnes 

Timnit Gebru

https://research.google/pubs/pub48823/

Closing the AI accountability gap: 

defining an end-to-end framework for internal algorithmic auditing.

in FAT* Barcelona, 2020, 

(ACM Conference on Fairness, Accountability, and Transparency)


https://scholar.archive.org/work/txsatfmnp5cdjbuowpt6j773ia/access/wayback/https://dl.acm.org/doi/pdf/10.1145/3351095.3372873

Rising concern for the societal implications of

artificial intelligence systems 

has inspired a wave of academic and journalistic literature

in which deployed systems are audited for harm

by investigatorsfrom outside the organizations deploying the algorithms. 

However, it remains challenging for practitioners to identify 

the harmful repercussions of their own systems 

prior to deployment, 

and, once deployed, emergent issues can become 

difficult or impossible to trace back to their source.

In this paper, we introduce a framework for

algorithmic auditing that supports 

artificial intelligence system development 

end-to-end, to be applied throughout the 

internal organization development life-cycle. 

Each stage of the audit yields a set of documents 

that together form an overall audit report, 

drawing on an organization’s values or principles

to assess the fit of decisions made throughout the process. 

The proposed auditing framework is intended to contribute to 

closing the accountability gap 

in the development and deployment of

large-scale artificial intelligence systems

by embedding a robust process to ensure audit integrity.


2018:

Margaret Mitchell,

Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, 

Timnit Gebru 

https://arxiv.org/abs/1810.03993

Model Cards for Model Reporting.


https://research.latinxinai.org/papers/neurips/2018/pdf/Oral_Andrew_Zaldivar.pdf

Trained machine learning models are increasingly used to perform 

high-impact tasks in areas such as law enforcement, 

medicine, education, and employment. 

In order to clarify the intended use cases of machine learning models 

and minimize their usage in contexts for which they are

not well suited, we recommend that

released models be accompanied by documentation 

detailing their performance characteristics. 

In this paper, we propose a framework that we call

model cards, 

to encourage such transparent model reporting. 

Model cards are short documents accompanying trained machine learning models

that provide benchmarked evaluation in a variety of conditions, 

such as across different cultural, demographic, or phenotypic groups

(e.g., race, geographic location, sex, Fitzpatrick skin type) 

and intersectional groups

(e.g., age and race, or sex and Fitzpatrick skin type) 

that are relevant to the intended application domains. 

Model cards also disclose the context in which

models are intended to be used,

details of the performance evaluation procedures, 

and other relevant information. 

While we focus primarily on human-centered machine learning models 

in the application fields of computer vision and natural language processing, 

this framework can be used to document 

any trained machine learning model. 

To solidify the concept, we provide cards for two 

supervised models: 

One trained to detect smiling faces in images, 

and one trained to detect toxic comments in text. 

We propose model cards as a step towards the 

responsible democratization of machine learning

and related AI technology, 

increasing transparency into how well AI technology works. 

We hope this work encourages those 

releasing trained machine learning models

to accompany model releases with 

similar detailed evaluation numbers

and other relevant documentation.


No comments:

Post a Comment