Harini Suresh @harini824 Twitter profile

Last Seen Profiles

@folsdon

@SoutheastOrthos

@SkribzSt

@DocPeteChambers

@OCZekip

@Fidelesmusic

@TiredAshleyGay

@Nady_clt

@Climatehope2

@clasicxs

@ante_vii

@GUEOPIC

@HKubbins

@Strainchamps

@BethHaven2

@VenezuelanVixen

@celesse

@ByromRaeghan

@apd_tx

@Bengaligoddess

@sharondixo73177

@danielfeller_ft

@ML_PhDer

@ging_air

@sadie_black10

@Marcus_colerick

@mblank

@EarthAngel11228

@damonibk

@_doe_____

@TaylorBlakeNyc

@LCTedesco

@JKelleyK

@MRFS_dq10

@__s0r__

@el_bonga

Harini Suresh

@harini824

8 months

Some news! I'll be joining @BrownCSDept as an assistant professor in fall 2024 :) I'm excited to continue work on participatory, equitable & accountable ML✨

125

83

4K

Harini Suresh

@harini824

5 months

a reminder that I'm recruiting PhD students for Fall 2024 at Brown CS, and applications are due Dec. 15 :) If you're interested in participatory ML, public agency & accountability, identity representations & intersectionality, or related areas - consider applying!

Harini Suresh

@harini824

8 months

Some news! I'll be joining @BrownCSDept as an assistant professor in fall 2024 :) I'm excited to continue work on participatory, equitable & accountable ML✨

125

83

4K

18

286

1K

Harini Suresh

@harini824

1 year

While it still feels (very) surreal... I defended & turned in my thesis last week!

15

512

Harini Suresh

@harini824

3 years

Who is ML interpretability meant for, and how can we design methods + interfaces that are truly understandable and relevant for them? Our CHI paper proposes a framework for thinking about these stakeholders beyond categorizations like “domain expert” or “model builder.”

2

36

228

Harini Suresh

@harini824

3 years

read @rajiinio 's thread! also: the updated version of this paper/fig includes "deployment bias," or harm arising from mistrust and/or misuse, which is a whole different beast from training data issues

Deb Raji

@rajiinio

3 years

1. Bias can start anywhere in the system - pre-processing, post-processing, with task design, with modeling choices, etc., in addition to issues with the data. The system arrives as the result of a lot of decisions, and any of those decisions can result in a biased outcome.

8

117

807

1

15

76

Harini Suresh

@harini824

3 years

we can't divorce "an algorithm" or "the data" from the systems & people that created it, and those who will use & be affected by it -- it's about power y'all! bumping this thread of a few papers/books I like that delve deeper into this:

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel @random_walker Fairness and Abstraction in Sociotechnical Systems () by @aselbst @zephoria @kdphd @geomblog @cyberlyra (2/5)

1

0

7

0

26

65

Harini Suresh

@harini824

8 months

Also, I am recruiting students for next year! If you are a prospective PhD student and think we may have overlapping research interests (or, if you know someone who is) - reach out!!

7

12

69

Harini Suresh

@harini824

8 months

@cornell_tech @2plus2make5 @karen_ec_levy Thanks for this writeup @BrownCSDept :)

Brown CS

@BrownCSDept

8 months

Please welcome @harini824 , who joins @BrownCSDept as Assistant Professor next fall! Read a full interview with her at Brown CS News:

0

3

31

1

3

60

Harini Suresh

@harini824

1 year

feeling lucky that I get to hang around these lab mates a little longer :')

MIT Clinical and Applied Machine Learning

@mit_caml

1 year

So proud of Harini Suresh ( @harini824 ) for (successfully!) defending a tour de force of a thesis on context and participation in machine learning. Pictured here: CAML reciprocates the immeasurable support she's given the lab over the course of her PhD 🤩🎉

0

5

36

8

2

57

Harini Suresh

@harini824

1 year

I'm presenting Kaleidoscope, a system for user-driven, context-specific, and semantically-meaningful ML model evaluation at #CHI2023 !

Kaleidoscope: Semantically-grounded, context-specific ML model evaluation | Proceedings of the 2023...

dl.acm.org

1

6

48

Harini Suresh

@harini824

8 months

Until then, I'll be doing a postdoc at @cornell_tech in NYC, working with @2plus2make5 @karen_ec_levy & Jon Kleinberg 🌆

2

0

44

Harini Suresh

@harini824

2 years

I'll be at @FAccTConference this week presenting "Towards Intersectional Feminist and Participatory ML: A Case Study in Supporting Feminicide Counterdata Collection"!

Towards Intersectional Feminist and Participatory ML: A Case Study in Supporting Feminicide...

dl.acm.org

Catherine D'Ignazio @[email protected]

@kanarinka

2 years

First, "Towards Intersectional Feminist and Participatory ML: A Case Study in Supporting Feminicide Counterdata Collection" comes out of the Data + Feminism Lab's collaboration with @ildalatam and @feminicidioURY 2/n

1

6

31

1

39

Harini Suresh

@harini824

5 years

The Problem with “Biased Data”

Sometimes when I tell people I research machine learning and it’s societal consequences, I get enthusiastic nods followed up by “right…

link.medium.com

2

8

32

Harini Suresh

@harini824

4 years

Thanks, @MD4SG ! I’m continually inspired by awesome collaborators @kanarinka @faeriedevilish @wonyoungso @ladelentes @ildalatam @WomenCountUSA Isa & Angeles :)

EAAMO

@EAAMO_ORG

4 years

New Horizons Award for the most inspiring works goes to: Feminicide & Machine Learning: Detecting Gender-based Violence to Strengthen Civil Sector Activism and Modeling Assumptions Clash with the Real World: Configuring Student Assignment Algorithms to Serve Community Needs

2

6

13

0

4

24

Harini Suresh

@harini824

5 years

really excited to share the first of many videos that @natalielaomit and I have been making. our goal is to make ML topics accessible and engaging, enable everyone to participate in public discussions & introduce responsible practices from the start

mltidbits

@mltidbits

5 years

Check out our first two videos on the #machinelearning lifecycle and different types of ML!

3

19

48

1

7

23

Harini Suresh

@harini824

1 year

Feeling grateful for *many* people including my thesis committee ( @arvindsatya1 @kanarinka and John Guttag), the support of 3 (!) amazing labs ( @mit_caml @mitvis and the data + feminism lab), my collaborators, friends, family, etc ❤️

2

0

16

Harini Suresh

@harini824

1 year

I'll be around MIT for the next few months & on the job market! please reach out if you know of an opportunity that would fit my background :)

3

9

16

Harini Suresh

@harini824

11 months

I’m at #FAccT2023 this week! send me your paper recs and/or let’s meet up :)

0

1

13

Harini Suresh

@harini824

5 years

almost makes me want to retake some (now-responsible) CS classes :)

Mozilla

@mozilla

5 years

Today, we're announcing the first winners of the #ResponsibleCS Challenge. @OmidyarNetwork , @mozilla , @SchmidtFutures & @craignewmark Philanthropies are awarding $2.4 million to 17 initiatives that integrate ethics into undergraduate computer science courses.

4

34

110

1

0

11

Harini Suresh

@harini824

3 years

@math_rachel Virginia Eubanks’ ( @PopTechWorks ) book has some great examples, like systems that make welfare decisions or allocate public housing

1

0

10

Harini Suresh

@harini824

4 years

One of the examples I found most disturbing from @joana_varon 's awesome talk was an algorithmic system deployed in Argentina, with the purpose of predicting a teenage girl's risk of pregnancy from her name and address (1/3)

1

2

11

Harini Suresh

@harini824

5 years

p.s. more detail in the paper:

A Framework for Understanding Sources of Harm throughout the...

As machine learning (ML) increasingly affects people and society, awareness of its potential unwanted consequences has also grown. To anticipate, prevent, and mitigate undesirable downstream...

arxiv.org

1

0

7

Harini Suresh

@harini824

5 years

@math_rachel glad you found our paper! feel free to send me a message if you have thoughts or comments :)

1

0

9

Harini Suresh

@harini824

3 years

For now, find it here! Shout out to wonderful collaborators @arvindsatya1 @steveg_cs @kevnam

1

0

9

Harini Suresh

@harini824

4 years

@joana_varon Paz Peña & @joana_varon talk about these systems (and the ideology around them) through the lens of decolonization: (pg 8-12, and also the whole issue is interesting) (3/3)

0

1

8

Harini Suresh

@harini824

4 years

@joana_varon There are so many messed up things about this, from the framing of teenage pregnancy as "predestined", to the way that pitching "big data" as the solution to complicated social problems detracts from actual policy reform (2/3)

1

0

7

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel @random_walker Fairness and Abstraction in Sociotechnical Systems () by @aselbst @zephoria @kdphd @geomblog @cyberlyra (2/5)

1

0

7

Harini Suresh

@harini824

4 years

i feel seen

Mary Louise Kelly

@NPRKelly

4 years

Anyone else in their kitchen sipping red wine and aggressively baking banana bread at 9:40pm? No? Just me? #coronavirusbaking

1K

271

11K

0

7

Harini Suresh

@harini824

3 years

and can also reveal users who are currently underserved: e.g., people w/o traditional expertise often have valuable personal knowledge + lived experiences that make them well-suited to catch errors, but debugging tools are primarily designed for people w/ formal ML knowledge.

1

2

7

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel @random_walker @aselbst @zephoria @kdphd @geomblog @cyberlyra Data Science as Political Action () by @benzevgreen (3/5)

1

0

6

Harini Suresh

@harini824

3 years

The more granular breakdown of expertise also helps reveal the heterogeneity in categorizations like "non-expert." Can we instead think about what effective interpretability looks like for people w/ rich personal or instrumental knowledge in areas beyond ML?

1

6

Harini Suresh

@harini824

3 years

Finally, we hope it can inspire designs that are built for previously overlooked users, and that more explicitly characterizing stakeholders can facilitate reflexivity about who is/is not included in design processes. More details in the paper :)

1

6

Harini Suresh

@harini824

5 years

apt airport TV on the way to #fat2019

0

6

Harini Suresh

@harini824

2 years

Our paper asks how we can meaningfully apply intersectional feminist + participatory methods *from the start*, to conceptualize + design ML tools that center and aim to challenge power inequalities.

1

0

5

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel @random_walker I would read it :) but I'd also check out these works by people who look at the problem from different perspectives than me: (1/5)

1

0

5

Harini Suresh

@harini824

2 years

Work w/ amazing Data + Feminism collaborators @kanarinka @rajivmovva @Angemartinezcub @cruxen @rahulbot @giulia_taurino @wonyoungso & Amelia Dogan

1

5

Harini Suresh

@harini824

3 years

Looking forward, we imagine the framework being used to identify and characterize a broader range of ML stakeholders and their interpretability needs. We also think it can make designing and generalizing from interpretability user studies more straightforward.

1

5

Harini Suresh

@harini824

2 years

I'll also be answering Q's with the wonderful @willieboag and @BiancaLepe on our paper about tech worker organizing as a lever for change!

Tech Worker Organizing for Power and Accountability | Proceedings of the 2022 ACM Conference on...

dl.acm.org

Catherine D'Ignazio @[email protected]

@kanarinka

2 years

In it, we review collective actions in the tech industry from the CAIT database - - and explore case studies of successful tech worker campaigns. 14/n

2

1

11

0

5

Harini Suresh

@harini824

3 years

note: We view "interpretability" broadly, i.e. any (static or interactive) insight into the system, whether prediction uncertainty, data collection info, more mechanistic explanations, etc. We use “interpretability” because of precedent but welcome ideas re: better/broader terms.

1

4

Harini Suresh

@harini824

3 years

In the paper, we pull from different fields to 1) expand "expertise" into different *types* of knowledge with different *contexts* and 2) distill a multi-level typology of interpretability needs that spans overarching goals, shorter-term objectives, and immediate tasks.

1

4

Harini Suresh

@harini824

5 years

can someone explain why “multi for her” vitamins are $2 more than “multi for him” @NatureMade @Walgreens

3

2

4

Harini Suresh

@harini824

3 years

Okay well, the website is down. But you can also find it on arXiv:

Beyond Expertise and Roles: A Framework to Characterize the...

To ensure accountability and mitigate harm, it is critical that diverse stakeholders can interrogate black-box automated systems and find information that is understandable, relevant, and useful...

arxiv.org

0

4

Harini Suresh

@harini824

1 year

We compared Kaleidoscope to other methods of grouping examples to test models, and did a user study w/ Reddit users & moderators - check out the details at

1

0

2

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel Studying up: reorienting the study of algorithmic fairness around issues of power () by @chels_bar et al (4/5)

Studying up | Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency

dl.acm.org

1

0

3

Harini Suresh

@harini824

3 years

Importantly, the framework decouples someone's knowledge from their needs: e.g., acting on model outputs is relevant beyond "decision-makers," debugging is impt not only for "ML experts," etc. This creates a more composable language for discovering & characterizing stakeholders

1

3

Harini Suresh

@harini824

2 years

We investigate this question through the lens of a case study: co-designing datasets and models to support the efforts of activists who collect and monitor counterdata (data that's currently neglected by mainstream institutions) about gender-related violence.

1

0

3

Harini Suresh

@harini824

4 years

@jaspervandorp @math_rachel Data Feminism () by @kanarinka @laurenfklein (5/5)

0

3

Harini Suresh

@harini824

5 years

@KLdivergence Thank you Kristian! :) I fixed it to point to a slightly updated version that's also on arxiv ()

A Framework for Understanding Sources of Harm throughout the...

As machine learning (ML) increasingly affects people and society, awareness of its potential unwanted consequences has also grown. To anticipate, prevent, and mitigate undesirable downstream...

arxiv.org

0

2

Harini Suresh

@harini824

1 year

i'm presenting bright & early at 9am on Thursday - hope to see you! and to anyone in Hamburg, I'd love to meet up :)

0

1

Harini Suresh

@harini824

5 years

@m_deff @math_rachel @jovialjoy @jeremyphoward @seb_ruder I think papers are important for people who want to dig in and build on your work; but media like blog posts are a way to reach people outside of your specific field to whom the paper may not be accessible

0

2

Harini Suresh

@harini824

2 years

The paper describes our approach to this case study in detail, and we also hope it can provide inspiration for how to mobilize intersectional feminist values in tech more broadly. P.S. would love to meet up in Seoul if you are here :')

1

0

2

Harini Suresh

@harini824

5 years

@jovialjoy This is a really good point, and definitely worth including. Thank you for the feedback!

0

1

Harini Suresh

@harini824

8 months

@brianwilt Aww thank you Brian!! I have lots of fond memories of that internship :D

1

0

1

Harini Suresh

@harini824

1 year

With the interactive system, users identify examples of important concepts, generalize them into larger, representative sets, and specify and test model behavior with them in semantically-meaningful ways

1

0

Harini Suresh

@harini824

5 years

@JordanBHarrod Interested :)

1

0

1

Harini Suresh

@harini824

8 months

@arvindsatya1 @BrownCSDept Thank you Arvind! Very grateful for all your guidance & support :D

0

1

Harini Suresh

@harini824

1 year

Evaluation/auditing is super important -- it guides what systems are considered state-of-the-art and deployed, and shapes broader research agendas/values in ML

1

0

Harini Suresh

@harini824

1 year

But existing evaluation methods/datasets don't really capture how the kinds of things that are important to people differ widely across contexts & communities, and might change over time

1

0

Harini Suresh

@harini824

1 year

Kaleidoscope is an alternative paradigm for model evaluation that aims to fill this gap, allowing users with deep personal knowledge of their context to translate their implicit expectations of "good model behavior" into concrete example sets and tests

1

0

Harini Suresh

@harini824

5 years

@Aaron_Horowitz @yonashav i’m mostly worried about putting the onus on individuals to “fix themselves” when larger issues like poverty or access to education/healthcare might motivate different interventions that are de-prioritized in algorithms that codify an individualistic viewpoint

0

1

Harini Suresh

@harini824

1 year

As a concrete example, it'd be really hard for a moderator of an online forum to ask a Q like "how well does this moderation model flag comments mocking someone's appearance?" -- this involves understanding a high-level concept stemming from the user's personal knowledge

1

0

Harini Suresh

@harini824

4 years

@Nale Yay, glad you found it! :D

1

0

1

Harini Suresh

@harini824

6 years

@irenetrampoline i asked nicole this after the talk - like @berkustun said, in the paper there are more details about the transfer learning they did b/w groups. in my experience the success of that type of transfer learning is tied to how similar the groups are

1

0

1

Harini Suresh

@harini824

2 years

@aspenkhopkins 🥰

0

1

Harini Suresh

@harini824

5 years

@EliasWalyBa Aww thanks! I’m glad you found it useful :D

0