Hidden Stories — an investigation of the stories that are being told about the refugee crisis
TLDR: A few perspectives from Cybernetics; Stories, Systems and Change; Representations in Social Science vs. Machine Learning; Minority report detection in news related to the refugee crisis; The power of community-driven investigative journalism; Seeing whole systems; [Part 1]
Reveal rather than enframe
The worldview of Cybernetics was about systems — human, nonhuman, and both — and that our world of “exceedingly complex systems”[1] can never be fully knowable, it’s instead a “performative dance of agency”[2]. Norbert Wiener talked about the coordination and collaboration between men and machines at the very core of Cybernetics. In his book “God and Golemn”, 1964, he wrote about many of the problems we have today in the way technology interacts with other systems. The “gadget worshipers”, he writes, “have the special motive of the desire to avoid personal responsibility for a dangerous or disastrous decision by placing the responsibility elsewhere”.
Cybernetics has had to address the critique of being seen as “a science of control”. In his book “The Cybernetic Brain: Sketches of Another Future”, 2010, Andrew Pickering makes the distinction between two senses of “control”. First, there’s the hierarchical, linear “command and control” type of “a power that flows in just one direction in the form of instructions for action (from one group of people to another, or, less conventionally, from humans to matter)”. His amazing perspective get’s us to think about the question of power within society but also in our relationship with Nature. He goes on to argue that “the cybernetics sense of control was not like that. Instead, in line with its ontology of unknowability and becoming, the cybernetic sense of control was rather one of getting along with, coping with, even taking advantage of an enjoying, a world that one cannot push around in that way. Even in its most asymmetric early moments, cybernetics never imagined that the classical mode of control was in fact possible”. Cybernetics demands a performative form of democracy where we use technology to reveal rather than enframe.
Stories and Systems
We’re woven into a web a personal and cultural stories. I somehow don’t laugh when a friend jokingly says something about communism as probably the story of communism in my head is different than other people’s image of it and that is all perfectly normal. For example, many times when we see a number our mind is quick to create many different stories about it. This has become increasingly easy with the help of technology, Statistics and AI in particular. In her essay “Leverage Points: Places to Intervene in a System”, Donella Meadows puts numbers in the very bottom of the list. Leverage points are “places within a complex system (a corporation, an economy, a living body, a city, an ecosystem) where a small shift in one thing can produce big changes in everything”. In her analysis numbers make up the least effective leverage point in a System. Here’s the full list:
PLACES TO INTERVENE IN A SYSTEM
(in increasing order of effectiveness)12. Constants, parameters, numbers (such as subsidies, taxes, standards).
11. The sizes of buffers and other stabilizing stocks, relative to their flows.
10. The structure of material stocks and flows (such as transport networks, population age structures).
9. The lengths of delays, relative to the rate of system change.
8. The strength of negative feedback loops, relative to the impacts they are trying to correct against.
7. The gain around driving positive feedback loops.
6. The structure of information flows (who does and does not have access to information).
5. The rules of the system (such as incentives, punishments, constraints).
4. The power to add, change, evolve, or self-organize system structure.
3. The goals of the system.
2. The mindset or paradigm out of which the system — its goals, structure, rules, delays, parameters — arises.
1. The power to transcend paradigms.
I think it’s interesting to notice that the more effective the leverage point, the more it has to do with stories. What’s the goal of a system, where does it come from? What’s a paradigm if not a story, a pattern, a worldview underlying the theories and methodology of a particular scientific subject. If we are to transcend paradigms and realise that no one paradigm is the ultimate truth — then numbers are absolutely needed but perhaps not the only thing to focus on.
Representations as part of our human condition
A story is intimately connected to the concept of representation. Social Science talks about the different meanings of a representation. First we can distinguish between the act or action of representing and the state of being represented. In the first case we are creating meaning by actively picturing, describing in language or in some other form. In the latter case the representation literally takes the place of the subject of our representation in some respect.
Philosophers René Descartes and John Locke saw mental representations as a mirror of objective reality. Furthermore we can distinguish between individual and collective representations, introduced by the sociologist Emile Durkheim, 1898. All of these clusters of meaning are critically important to take into account as they show how a story literally re-presents something or someone.
Kate Crawford from the AI Now Institute, in her keynote during the NIPS conference in 2017 talked about the social and economic implications of bias in Machine Learning systems. Specifically, investigating if algorithms have negative externalities such as allocative and representational harms. Harms of allocation are often related to numbers — groups or individuals are denied access to some kind of resources or opportunities. For example, the denial of mortgages to people who live within a particular zip code or using algorithmic scores to decide who’s more likely to do better on the job[3] or even who deserves to see the advertisement about it.
Harms of representation are related to the way a system may unintentionally underscore or reinforce the subordination of some social and cultural groups.Many times, these harms are much more related to stories rather than numbers and this makes them extremely hard to analyze. “The perspective of representational harms requires us to move beyond biases in the data set and think about the role of ML in harmful representations of human identity, and how these biases reinforce the subordination of groups along the lines of identity and affect how groups or individuals are understood socially, thereby also contributing to harmful attitudes and cultural beliefs in the longer term.” — to quote great work by researchers in this field[4]. In their proposal of Algorithmic Impact Assessments, the AI Now Institute urge us to consider both of these kinds of harms in our interactions with technology.
Why do we want to learn representations when doing Machine Learning?
In their book on Deep Learning, Ian Goodfellow, Yoshua Bengio, and Aaaron Courville ask the question about what makes one representation better than another? An ideal representation, they say, “is one that disentangles the underlying causal factors of variation that generated the data”. We describe how data is generated by using probability distributions. Many deep learning algorithms are motivated by the assumption that the hidden units of neural networks can learn to represent the underlying causal factors that explain the data. This hypothesis is still a huge challenge and opportunity for us to critique the algorithms we create[5].
In doing research we use deterministic feedforward networks as universal approximators of functions. Many of the so called structured probabilistic models with a single hidden layer of latent variables, including restricted Boltzmann machines and deep belief networks, are universal approximators of probability distributions. Ultimately I think that learning representations in Machine Learning gives us new perspectives about hidden relationships.
Social Representation theory is part of the field of social psychology which studies formation and transformation of meanings, knowledge, beliefs, and actions of complex social phenomena like democracy, human rights, or mental illness, in and through communication and culture. Following the interdisciplinary approach, I think researchers in the field of Social Representation are already looking at technology as another key contributing factor. On the other hand, the least that technologists could learn from the field of social psychology is that representations are dynamic, performative, inherently partial as well as an active process and product of power.
We need to be thinking about representations that go beyond a specific ML algorithm and are informed by the broader context where the algorithm is going to be deployed at. These algorithmic representation need to be in feedback loops with the other contributing factors to Social Representations — communication, culture, and others, and therefore they need to be interactive and adaptable.
Aligned with this mindset I think that algorithmic tools could aid multilingual cross-cultural understanding if we could create a broader conversation where vulnerable communities are empowered to participate. This is the focus of an investigation into the stories that are being told about the refugee crisis. I’ve been so glad to be part of this collaboration between Nick De Palma, Brent Dixon and myself. By working on this project, we set out to understand the challenges facing those fleeing violence, turmoil, or seeking opportunity in a foreign country where they are often denied their basic human rights and liberties. Learning from Cybernetics, we wanted to understand if an algorithmic system can be used to reveal hidden relationships and anomalies in a database of stories related to the refugee-crisis. We were glad to have the opportunity to learn from the amazing team at Greece Communitere — the Greece chapter of an international NGO creating dynamic, collaborative hubs in displaced and post-disaster communities.
Investigative community-driven journalism within refugee settlements could reduce report fatigue while allowing a machine learning system to aid the work of organizations and individuals where the algorithmic system is used for intervention rather than prediction. Rather than build tools to aggregate data, we hope to use AI techniques to reveal patterns and highlight minority reports that need attention and resources.
- Investigative: What is investigative journalism? The best definition of what constitutes “investigative” journalism is journalism that seeks to reveal something that someone with some level of power (a person, group or institution) seeks to keep a secret[6].
- Community-driven: Communities bring the energy and expertise to reinvent themselves from within. We wanted to investigate and learn from the success stories of how civic journalism used to serve communities in the 1990s. The focus of civic journalism was to go beyond informing the public and help provide support and interventions that enriched community life. Its founders described it as “a craft that builds up the world while simultaneously describing it”[7].
- Journalism: Journalism has historically played a central role in shaping public discourse. We wanted to understand the challenges journalists face in their invaluable work to help some of the most vulnerable people in the world in the setting of refugee camps.
- AI tools for intervention rather than prediction: we wanted to research what kinds of technological tools are already in place in the broader ecosystem of organizations and individuals working to provide sustainable solutions to the problems facing displaced communities.
The Dataset
The goal of our work was to analyze if AI-enabled topic modeling and anomaly detection can aid the work of humanitarian action and advocacy groups, and investigative journalists by helping them organize different pieces of evidence. We begin by acknowledging that a transparent data collection process is crucial for the real-world success of any potential proposal. It must preserve people’s privacy and must be informed by all involved stakeholders. For the purpose of our experiments, we use a database consisting of 6,258 news articles published by major media outlets between 01/2016–09/2017 and mentioning the refugee crisis[8]. However, we hope in future work to collaborate directly with practitioners to help them extract and map insights from case work related to human rights violations in refugee settlements.
We used an unsupervised learning algorithm to learn representations and then cluster similar representations together. In the image above you see all the news articles in the corpus, where each one of them is represented by a token whose color depends on the cluster it is part of. By analyzing word frequencies in each cluster we find that for more than half of all clusters the US President Trump is a major actor, however individual subclusters were formed related to other political figures such as the US Secretary of State John Kerry and Germany’s Chancellor Angela Merkel. Another subcluster emerges related to terms such as women, public health and policy. Finally, we use the representations to extract the distance between an article and its reconstruction in a high dimensional space. Our main hypothesis in the work is that some of the articles were not modeled well and that their reconstruction error would be significantly larger than the other articles in the news corpus.
Initial probing of these anomalous news reports show that they cover the meta-topics such as: 1) articles not related to the refugee crisis at all, 2) articles expressing specific sentiment which we don’t see often in media, 3) articles which were very concrete and graphic about violence, 4) articles about the Cold War, and 5) an assortment of other topics.
We are very glad and excited to have the opportunity to present this work during the upcoming AI For Social Good workshop at the NIPS Machine Learning conference.
What becomes possible if we have good representations?
Seeing Whole Systems
The image above is from a report released by NASA on September 26th visualizing how Hurricane Florence impacted the Carolina coastline as polluted rivers dump into the Atlantic.
Pictures are representations. In his work Picture Theory, W. J. T. Mitchell points out that as representations they not only work to mediate our “knowledge (of slavery and of many other things), but obstructs, fragments, and negates that knowledge”. The access to satellite imaging has a huge multidisciplinary impact on all levels and gives us an opportunity to look for answers but even more so to ask better questions. I think that investigating the implications of AI-driven systems is an area where much more work is needed. Ultimately, by questioning the representations we are faced with, we could co-create a world where we promote peaceful and inclusive societies for sustainable development, provide access to justice for all and build effective, accountable and inclusive institutions at all levels[9].
References
[1] Stafford Beer’s concept of an exceedingly complex system — “a system with its own inner dynamics, with which we can interact, but which we can never exhaustively know, which can always surprise us”. The quote is from Andrew Pickering’s book “The Cybernetic Brain: Sketches of Another Future”.
[2] Andrew Pickering’s worldview about what it means to be within the world we live in — a world built from exceedingly complex systems.
[3] Barocas, Solon and Selbst, Andrew D., Big Data’s Disparate Impact (2016). 104 California Law Review 671 (2016). Available at SSRN: https://ssrn.com/abstract=2477899
[4] Dobbe, Roel, Dean, Sarah, Gilbert, Thomas and Kohli, Nitin, A Broader View on Bias in Automated Decision-Making: Reflecting on Epistemology and Dynamics https://arxiv.org/pdf/1807.00553.pdf
[5] Blei, David M., Build, Compute, Critique, Repeat: Data Analysis with Latent Variable Models
https://www.annualreviews.org/doi/abs/10.1146/annurev-statistics-022513-115657?journalCode=statistics
[6] Tofel, Richard J., A White Paper from ProPublica, Non-Profit Journalism: Issues Around Impact https://s3.amazonaws.com/propublica/assets/about/LFA_ProPublica-white-paper_2.1.pdf
[7] Overholser, Geneva, How to Best Serve Communities: Reflections on Civic Journalism
https://www.democracyfund.org/publications/how-to-best-serve-communities-reflections-on-civic-journalism
[8] A dataset available on Kaggle and containing 143,000 articles from 15 American publications https://www.kaggle.com/snapcrack/all-the-news
[9] UN Sustainable Development Goal #16 https://sustainabledevelopment.un.org/sdg16
