Retraining the machine: Addressing algorithmic bias

Digital Methods Summer School 2018


d16179728088_debbe41a19_q.jpg b5033431330_a9f9e6f176_q.jpg 7535026324_b001e1f6ee_q.jpg 7535043142_4c6997ddc6_q.jpg 7535058620_8d7c4b13b5_q.jpg 7535104948_bc96102eba_q.jpg


2–13 July 2018

Everyday location:
Media Studies
University of Amsterdam
1012 XT Amsterdam

This year’s Digital Methods Summer School is dedicated to approaches to studying so-called machine bias. Discussions have been focusing on how to hold algorithms accountable for discrimination in their outputting of results such as in the notorious cases of query results for ‘professional hair’ (white women’s hairdo’s) and ‘unprofessional hair’ (black womens’ hair-do’s). Recently, it was found that search engine image results for ‘pregnancy’ and ‘unwanted pregnancy’ are similarly divided, with the pregnancy queries returning white skinned women (mainly bellies, privileging the baby over the woman). ‘Unwanted pregnancy’ results in diverse ethnicities. These are new variations on classic, and still urgent, search engine critiques (once known as ‘googlearchies’) which questioned the hierarchies built into rankings, asking who is being authorised by the engine to provide the information. That work moves forward at the Summer School, building on examinations of the volatility of engine results, as in the Issue Dramaturg project, which put on display the drama of websites rising and falling in their rankings after algorithmic updates, meant to fight spam, but having unintended, epistemological consequences. More recently, Facebook newsfeeds have been the source of critique for their privileging and burying mechanisms, however much they — like the engine returns preceding them — are not easily captured and documented. Saving engine results has been against the terms of service; making derivative works out of engine results also breaks the user contract. Saving, or recording, social media (newsfeed) rolls seems even less practicable given how feeds are even more personalised, presumably resisting generalisable findings. User surveys pointing out unexpected newsfeed results have led to calls for ‘algorithmic auditing’, a precursor to machine bias critique. As reported in the technical press, querying social media ad interfaces shows highly segmented audiences (including racist ones such as publics to target for ‘jew haters’ among other available keyword audiences for sale).

These ad interface results could be repurposed to show which population segments (as defined by the platforms) are driving the content choices reflected in the results served. How large are these discriminatory segments? Capturing, auditing, or repurposing results are diagnostic practices, identifying under which circumstances machines could or ought to be retrained. The larger question, however, concerns how to retrain the machine. One approach lies in query design — fashioning queries so as to ‘re-bias’ the results. Others concern corpus development. For example in stock photography efforts have been to reimagine (‘re-image’) women (in the well-known case of the ‘lean in collection’), however much the images are often used out of context, as has been found. Yet another one concerns training and maturing research accounts to trigger controlled algorithmic responses.

The Digital Methods Summer School is interested in contributing not only to interpretations of celebrated cases of algorithmic or machine bias, but also providing diagnostic, query-related, research account and corpus-building research practices that seek to address the matter more conceptually.

Expanding the case study collection is also of interest; age discrimination in Facebook ad interfaces (an American theme) is a recent example of a telling case study of in-built rather than organic machine bias, but the international landscape may contribute more to bias detection, as is the aim of the Summer School. In Twitter there are feminist bots striving to keep the #metoo space serious, since the spam has arrived. Which other practices of remaining on topic may be found, and how may their success and and complications be characterised? There is also the question of the ramifications of conceptual contributions to re-biasing for big data science. Which practical contributions could be made to big data critique?

Summer School Philosophy

The Digital Methods Summer School is exploratory and experimental. It is not a setting for ‘just’ tool training or for principally tool-driven research. Substantive research projects are conceived and carried out. Participants are encouraged to ‘span time with their issue’ and the materials. In other words, we heed Alexander Galloway’s admonition about data and tool-driven work: “Those who were formerly scholars or experts in a certain area are now recast as mere tool users beholden to the affordances of the tool — while students spend ever more time mastering menus and buttons, becoming literate in a digital device rather than a literary corpus” (Galloway 2014:127). We encourage device and corpus literacy! The device training we ask you to do prior to the Summer School through online tutorials, and at the Summer School itself, in a kind of flipped learning environment (if you'll excuse the overused phrase), we would like to believe that you have familiarised yourself already with the tools and completed the tutorials available online. During the Summer School we will discuss and tinker with the nitty-gritty, aim to invent new methods, techniques and heuristics and create the first iterations of compelling work to be shared.

About Digital Methods as a Concept

Digital methods is a term coined as a counterpoint to virtual methods, which typically digitize existing methods and port them onto the Web. Digital methods, contrariwise, seek to learn from the methods built into the dominant devices online, and repurpose them for social and cultural research. That is, the challenge is to study both the info-web as well as the social web with the tools that organize them. There is a general protocol to digital methods. At the outset stock is taken of the natively digital objects that are available (links, tags, threads, etc.) and how devices such as search engines make use of them. Can the device techniques be repurposed, for example by remixing the digital objects they take as inputs? Once findings are made with online data, where to ground them? Is the baseline still the offline, or are findings to be grounded in more online data? Taking up these questions more theoretically (but also practically) there is also a Digital Methods book (MIT Press, 2013) as well as a complementary Issue Mapping book (Amsterdam University Press, 2015), and other digital methods publications.

About the Digital Methods Summer School

The Digital Methods Summer School, founded ten years ago, in 2007, together with the Digital Methods Initiative, is directed by Prof. Richard Rogers, Chair in New Media & Digital Culture and Department Chair at Media Studies, University of Amsterdam. The Summer School is one training opportunity provided by the Digital Methods Initiative (DMI). DMI also has a Winter School, which includes a mini-conference, where papers are presented and responded to. Winter School papers are often the result of Summer School projects. The Summer School is coordinated by PhD candidates in New Media at the University of Amsterdam, or affiliates. This year the coordinators are Sabine Niederer, Natalia Sánchez-Quérubin and Fernando van der Vlist. The Summer School has a technical staff as well as a design staff, drawn from the ranks of Density Design in Milan. The Summer School also relies on a technical infrastructure of some nine servers hosting tools and storing data, which recently (and intrepidly) moved to the cloud. In a culture of experimentation and skill-sharing, participants bring their laptops, learn method, undertake research projects, make reports, tools and graphics and write them up on the Digital Methods wiki. The Summer School concludes with final presentations. Often there are subject matter experts from non-governmental or other organizations who present their analytical needs and issues at the outset and the projects seek to meet those needs, however indirectly. For instance, Women on Wavescame along during the 2010, Fair Phone to the 2012 Summer School and Greenpeace International and their Gezi Park project in 2013 as well as the COP21 Lima project in 2015. We have worked on the issue of rewilding eco-spaces with NGOs in the 2014 Summer School. More recently we have sought to repopulate city dashboards (Summer School 2015 and Winter School 2017).

About the Digital Methods Initiative

The Digital Methods Summer School is part of the Digital Methods Initiative (DMI), Amsterdam, dedicated to developing methods for Internet-related research. DMI was founded a decade ago with a grant from the Mondriaan Foundation, and the Summer School has been supported by the Center for Creation, Content and Technology ( CCCT), University of Amsterdam, organized by the Faculty of Science with sponsorship from Platform Betatechniek. It also has received support from the Citizen Data Lab, Amsterdam University of Applied Sciences as well as "Media of Cooperation," University of Siegen.

Applications & Key Dates

To apply for the Digital Methods Summer School 2017, please use the University of Amsterdam Summer School form. If that form is not working well (as some applicants have reported), please send a one-page letter explaining how digital methods training would benefit your current work, and also enclose a CV (with full postal address), a copy of your passport (details page only), a headshot photo as well as a 100-word bio (to be included in the Summer School welcome package). Mark your application "DMI Training Certificate Program," and send to summerschool [at] digitalmethods.net.
  • The deadline for applications for the Summer School is 4 May 2017.
  • Notifications will be sent on 7 May. Accepted participants will receive a welcome package, which includes a reader, a schedule, and a face book of all participants.
  • The cost of the Summer School is eur 895 and is open to PhD candidates and motivated scholars as well as to research master’s students and advanced master’s students. Data journalists, artists, and research professionals are also welcome to apply. Accepted applicants will be informed of the bank transfer details upon notice of acceptance to the Summer School on 7 May.
  • The fee must be paid by 18 June. Participants must send a proof of payment by that date.
  • University of Amsterdam students are exempt from tuition and should state on the application form (under tuition fee remarks) that they wish to apply for a fee waiver. Please also provide your student number.
Any questions may be addressed to the Summer School coordinators Esther Weltevrede and Fernando van der Vlist: summerschool [at] digitalmethods.net. Informal queries may be sent to this email address as well.

Scholarships

The Digital Methods Summer School is part of the University of Amsterdam Summer School programme, which has a video giving a flavor of the Summer School experience. Students from universities in the LERU and U21 networks are eligible for a scholarship to help cover the cost of tuition for the DMI Summer School. Please state LERU or U21 university affiliation under tuition remarks when applying to the Summer School. Dutch universities are not eligible.

Accommodations & Catering

The Summer School is self-catered, and there are abundant cafes and a university mensa nearby. For a map we made of nearby lunch (and coffee) places, see bit.ly/dmi17_ss_lunch.

We have organised a discount at Hotel Casa Amsterdam. For those who prefer other accommodations, we suggest Airbnb or similar. For shorter stay, there is Hotel Le Coin, where you may request a university discount.

Successful Completion & Completion Certificates (incl. 6 ECTS when necessary)

To successfully complete the Summer School and receive a completion certificate (and 6 ECTS when necessary), you must complete a significant contribution to two Summer School projects (one in week one and the other in week two), evidenced by co-authorship of the project reports as well as final (joint) presentations. Templates for the project report as well as for the presentation slides are supplied.

Schedule

The Summer School meets every weekday. Please bring your laptop and your VGA adaptor for the projector. (An iPad is not enough.) We will provide connectivity. We start generally at 9:15 in the morning, and end around 17:30. There are morning talks as well as tutorials one to two days per week. All other time is devoted to project work with occasional collective and individual feedback sessions. Each Friday there are final presentations. On the second Friday we have a festive closing with a boat trip on the canals of Amsterdam.

Preparations: Online Tutorials

For your Summer School to be especially successful we would recommend highly that you watch (or listen to) the Digital Methods tutorials. The DMI YouTube channel has copious materials, and we would very much like for you to watch the social media tool tutorials.

Social Media & Participant Face Book

  • Twitter: #dmi18
  • We will have a list of Summer School participants and make an old-fashioned face book with the headshots and bios you send to us.

Evening Hangout Suggestions

Topic revision: r3 - 19 Feb 2018, RichardRogers
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Foswiki? Send feedback