Last week I attended the Australian Society of Archivists Annual Conference. One of my jobs was to talk about the cultural perspectives of technologies. I talked about algorithms, machine learning, and constructions of evidence of culture. I think I scared people.
Below is a video of the extended talk. It goes for just over 10 mins. At the end, I talk about my Mediated recordkeeping model and how it might be useful in exploring these expanding contexts and complexities of culture.
I am keen to explore the role of machine learning in cultural heritage spaces. Who wants to help?
Hello, I am Dr. Leisa Gibbons from Curtin University. I teach archives and preservation to undergrad and post grad students. In my research, I explore sociotechnical issues, impacts and implications of acquisition and preservation of online content and the role that archivists can, do and might play in the formation of digital cultural heritage.
In this presentation I am going to share with you some intriguing information about algorithms and machine learning I have been collecting over the last year or so, so that I might talk about the nature and purpose of web archiving and how it is possible to understand evidence of culture as it is being valued and formed over spacetime.
Originally, this presentation was designed in PechaKucha style where 20 slides are shown for 20 seconds each. This presentation has 13 slides with the last one being quite a lot longer than 20 seconds.
This year Professor Geoff Goodhill, from the University of Queensland wrote about AlphaGo, an AI program designed to learn to play Go. AlphaGo learns via use of neural networks and extraction of key ideas.
You’ve probably heard about the algorithm created by Standford researchers that predicts sexual orientation from photographs of a person’s face? This is also generated with learning neural network technology.
Yet, as Professor Geoff Goodhill mentioned, there is no known way to interrogate the network to directly read out what these key ideas are that help the algorithm make decisions. Instead they can only study its outputs and hope to learn from these.
A couple of years ago, Vladan Joler and colleagues in Belgrade began investigating the inner workings of Facebook. This image is a flow chart that they created on how our interactions with Facebook create data – which show how we, as Facebook users, are in fact doing unpaid work for Facebook – so they can sell us stuff.
We all know this of course, but perhaps we think less about what this might mean in 20 or 150 years time related to data privacy and surveillance when you think about the data we give Facebook is used to calculate our ethnic affinity (Facebook’s term), sexual orientation, political affiliation, social class, travel schedule and much more.
In 2013, a community of scholars and activists gathered in the US to examine and discuss the social justice impact of algorithmic accountability or #algacc. Tthey raised more questions than answers about the impact of data surveillance and our right to know what and how data collected about us is being used.
UCLA Assistant Professor Safiya Noble writes about algorithms of oppression and how the data they use to learn reinforces existing structures of racism and sexism. Safiya talks about how a Google search she undertook on the search term “black girls” often suggested porn sites and un-moderated discussions about “why black women are so sassy” or “why black women are so angry” – presenting a disturbing portrait of black womanhood in modern society.
Reseachers at the Pew Research Center identified seven main themes about the algorithm era.
As part of sharing these concerns they tell a story of how Microsoft engineers recently created a Twitter bot named “Tay” in an attempt to chat with Millennials by responding to their prompts, but within hours it was spouting racist, sexist, Holocaust-denying tweets based on algorithms that had it “learning” how to respond to others based on what was tweeted at it.
This year, US Professor Ben Shneiderman proposed that there should be a regulatory body called a National Algorithms Safety Board, which would provide oversight for high-stakes algorithms.
In Australia, there are at least 20 separate parts of law that allow the government to give computers the power to make decisions. Decisions that used to be made by a human and can have important consequences.
These laws allow for computers to make decisions about social security, taxation, parental leave, superannuation, migration, biosecurity and child support. In every case, some kind of algorithm may be used to make decisions, yet we have no knowledge of how these work.
These are powerful and disturbing stories about the creation and use of data, the role the internet plays and the shaping role that mathematics and computers are playing in our society. This brings me to web archiving.
One of the most basic tenants of all data science is that data doesn’t exist in a vacuum, it is the result of a massive pipeline of explicit and implicit decisions…
…yet so much of the output of the data science world proceeds as if data can be cleanly separated from the contexts in which it is created.
Nowhere is this more apparent than the world of web archiving.
Researcher Kalev Leetaru, wrote an article for Forbes recently that starts with this paragraph. This was not his first dig at how poorly web archiving is conceptualised and constructed. He started in 2012 talking about the lack of documentation regarding even the most critical decisions like inclusion criteria, seed lists and third-party crawl donors means that we have precious little insight into how these archives were constructed and what biases may be manifest through those myriad decisions.
This is not a new conversation for me either. But algorthms and the rate of change in our virtual spaces and technologies are raising the stakes.
When it comes to using data to understand the world around us, the most important question revolves around how well that data reflects the phenomena we are attempting to study.
Kalev rightly asks questions about the nature of web archiving. When it comes to using data to understand the world around us, the most important question revolves around how well that data reflects the phenomena we are attempting to study. Do Twitter-based studies of human society truly reflect the dreams and fears of global society or are they systematically biased geographically and demographically? Do the breaking news events surfaced by the Facebook Trending Topics module exclude much of the continent of Africa and is Africa as a whole largely absent from the datasets we use to understand the world? Does the relative dearth of analytic algorithms for languages other than English mean we miss critical trends.
All this exploration of algorithms and the internet comes back to a question I have been raising for a decade now – what is evidence of culture? And in this question, what is the role of the archivist and the archives in the construction and dissemination of cultural heritage?
If web archives are online cultural heritage, how is their construction being understood and documented? As Kalev points out – does the medium examined define the results?
This raises the question of what web archives actually evidence of? But how do we interrogate the notion of evidence of culture?
Image: Mediated recordkeeping model
I want to share with you a model I created from research on how to understand the complexity of evidence of culture in online spaces. This model is an attempt to make sense of how and why people interact with recorded information – the purposes, the values, and the nature of memory as it is created, shared, accessed and managed over time in various and complex ways, including in response to technologies, other people and entities, and various mechanisms, systems and tools that help to enable and empower, as well as disempower and make hidden.
I want to share with you the three important areas it represents:
Firstly, memory and evidence as processes are separate but intrinsically linked. The processes of memory-making has a relationship to multiple systems, mechanisms and perspectives involved in establishing evidence.
Secondly, how people create is linked to how they see and identify themselves, what they are interested in, how they identify with various communities, as well as what values they perceive according to various community cultures. Narrative is vital to understanding this as it is a tool that can construct and communicate multiple and simultaneous realities, identify and make sense of the self within groups, community and society, and is imbued with power; of dominant, counter and competing narratives and as a mechanism for memory-making and knowledge preservation.
Thirdly, interaction occurs in conjunction with an understanding of action at various levels, as well as in relation to how people use, value and experience technologies including what technologies afford or do not allow to help people achieve their goals in creating and sharing something of who they are.
This model shows all these points of view to exist simultaneously and in multiples. How an individual understands their identity and work is not necessarily how it is seen by someone else. So when the archivist creates, in the creator dimension by documenting the world, they should be taking into account the varied, diverse and potentially incommensurable complexities that make up this map of how we understand cultural heritage as evidence of culture.
If we see algorithms as part of a continuum of mediated memories where and how do they fit in? Whose narratives are being told and what do we need to know about mandates to understand their contexts as memory? I don’t have any answers today but this is something I am about to examine.
But what my research into algorithms is beginning to reveal is the deep complex relationship and nature we have with data and machines. Recordkeeping is a memory-making process that contributes to evolving values, purposes and interactions over spacetime including memory (as making and remembering), narrative (as personal, sharing and evolving), evidence (as constructions of value and meaning) and technologies (as mediators and facilitators).
Archivists, and I count myself as one, need to consider what this means as to how we understand culture as evidence and heritage as it is being formed. Archivists also needs to understand and challenge their role in the system so that they may empower, discover and transform to meet multiple needs over time. Flexibility, adaptability and a need to understand what is being valued and who by as it is being created is essential to any transformation. That includes transformation within ourselves as professionals as well as the transformation of what role archives as constructions of evidence play in society.