The Air Pump

Speculative diagrams: Experiments in mapping Youtube

2019-04-02T00:00:00Z

Marenko, B. and Benqué, D. (2019) ‘Speculative diagrams: Experiments in mapping Youtube’, in Method & Critique; frictions and shifts in RTD. Research Through Design, TU Delft. doi: https://doi.org/10.6084/m9.figshare.7855811.v1.

Betti Marenko and I presented a paper at the 2019 Research Through Design Conference: Method & Critique, frictions and shifts in RTD. This was a development of my previous work attempting to map recommendations on Youtube. I developed a new visualisation that focuses on individual traces, following individual paths through Youtube indefinitely. The paper is framed by our ongoing conversation linking design theory, critical practice, algorithmic prediction, and diagrams.

For more details see the paper and the Gitlab repository.

Speculative diagrams: plotting to reclaim algorithmic prediction

2019-01-17T00:00:00Z

The following presentation was given by Betti Marenko and David Benqué on the 3rd of June 2018 at Design Anthropology: Uniting experience and imagination in the midst of social and material transformation, a panel of the Art, Materiality and Representation conference organised by the Royal Anthropological Institute at the British Museum in London.

How can this proposal be distinguished from issues of authority and generality currently articulated to the notion of “theory”? This question is particularly important since the “cosmopolitical” proposal, as I intend to characterize it, is not designed primarily for “generalists”; it has meaning only in concrete situations where practitioners operate. It furthermore requires practitioners who (and this is a political problem, not a cosmopolitical one) have learned to shrug their shoulders at the claims of generalizing theoreticians that define them as subordinates charged with the task of “applying” a theory or that capture their practice as an illustration of a theory. --- Isabelle Stengers (2005)

We begin from Isabelle Stengers’ quotation as an opening gambit to frame a collaborative practice that brings together a theorist and a designer with a common interest in computation, speculation and divination as a way to un-settle algorithmic determinism. We chose to work with diagrams to figure out how to reclaim the spaces of potential from the clutches of algorithmic prediction and its foreclosure of futures.

Arithmetic 12 Chart, William Delbert Gann, circa 1935.

We focus on diagrams because they are a common language between computation, prediction and speculation.

Wu, X. and Zhang, X. (2016) ‘Automated Inference on Criminality using Face Images’, arXiv.org.

On one hand diagrams are the operational core of algorithmic prediction, they are pattern-finding machines that produce the vector spaces underpinning the current regime of governmentality .

(Deleuze and Guattari, 1987)

On the other hand, however, diagrams are speculative visualisations of what cannot be seen, yet; inventive machines that map the unformed and the unstable as they feed into change.

Hence, the lure of diagram-making, for us, is to delve into the possible as a fluid rather than a solid material in a way that retains openness to the unknown.

We insist on the creative dimension of diagram-making as a strategy to counteract what Félix Guattari describes as the drive to “binarise the possible incessantly, close off the future through all sorts of procedures” (Guattari 2016, p.192). A way in which the possible is binarised and the future is foreclosed right now is through algorithmic governmentality.

Algorithmic governmentality is the contemporary pre-emptive regime of power that “bypasses consciousness and reflexivity, and operates on the mode of alerts and reflexes” (Rouvroy, 2013, p. 153).

Classifier comparison, Scikit Learn (Pedregosa et al., 2011)

Algorithmic governmentality operates through Machine Learning algorithms performing continuous data-trawling, autonomous learning, recursive training and re-modelling in ‘real-time’. Through clustering, classifying, categorizing, and matching, Machine Learning predicts future behaviours based on past occurrences.

This is where prediction becomes prescription: a meta-digital phase (Parisi, 2017) where the constant supply of fresh data ensures a continuous automatic revision and refinement of models.

Wright, Joe. Black Mirror S03E01 Nosedive, 2016.

We can say, then, that the real target of algorithmic governmentality is the “inactual, potential dimensions of human existence, its dimensions of virtuality, the conditional mode of what people ‘could’ do, their potency or agency” (Rouvroy, 2013).

What is design doing about this?

Our argument is a critical one. We critique design for failing to engage with algorithmic prediction outside of mere implementation and ‘anticipatory design’ a user-experience approach which seeks to design seamlessly for a future without choice.

Even Speculative and critical design seems unable to address the practices and politics of algorithmic governmentality because of its narrow understanding of ‘speculation’.

An example of this is given by a pervasive Speculative and critical design trope:
Stuart Candy’s Futures Cone (2010) used as a foundational diagram to describe the practice of Speculative and critical design since its inception. However, in a current moment saturated by computational prediction, this single ‘conceptual map’ fails to capture, and position itself in relation to, the continuous diagramming of the future by machines.

We think this is a problem.

Echoing Isabelle Stengers, this is a political problem: the Futures Cone, in its more mainstream and corporate dissemination, has been stripped bare of its political connotations, and in some cases co-opted in the rhetoric of innovation - we are specifically attentive to how certain aspects of the preferable may sit beyond the possible, an explicit reference to an activist, utopian, practical political engagement.

The preferable in the cone is where the politics are.

The preferable is the space of the maybe.

Our counter proposals are new diagrams that bring together an idea of computation as transformative, creative and inventive, engaging with ‘speculative gestures’ (Debaise and Stengers, 2015) as well as emphasizing the political valence of divinatory practices (Ramey, 2016) and conjectural knowledge (Ginzburg, 1980), to re-invigorate the speculative’ in Speculative and critical design.

Speculative gestures are necessary to counteract the current crisis in modes of thinking in which design is complicit - and to propose new modes of existence - with design as a powerful driver for change

We take divination as a practice that uses chance as an occasion to make meaning through a inquiry into the unknown; and making conjecture as a type of knowledge-production that allows for elements of unpredictability that are not measurable but are situated.

Making conjecture is also the practical open-ended nature of our speculative collaboration, in itself a diagrammable activity that requires invention, patience and lack of concern for the need to resolve differences. Which is our preferred notion of speculation - less future-fixed and less fixed on explaining, resolving and offering solutions, but instead a Speculative thinking/practice that aims at maximizing friction with experience.

The project here is to use computation as a picklock to pry open possible futures, pivoting on the potential, and staying with the indeterminate. These are some examples of our work in progress inspired by grooves, probeheads, holes, picklocks and string figures.

Grooves

When thinking gets stuck 'in grooves’, as Whitehead puts it, artifices and interventions are needed to 'activate' thinking, to cultivate power, to instigate lines of flight.

“Now to be mentally in a groove is to live in contemplating a given set of abstractions. The groove prevents straying across country, and the abstraction abstracts from something to which no further attention is paid. But there is no groove of abstractions which is adequate for the comprehension of human life” (Whitehead, 1970, p. 197)

Probe-heads

For Deleuze and Guattari (1987) probe-heads are “guiding devices that dismantle the strata in their wake, break through the walls of significance, pour out of the holes of subjectivity, fell trees in favour of veritable rhizomes, and steer flows down lines of positive deterritorialization or creative flight.” (Deleuze and Guattari, 1987, p. 190)

But they also produce other, stranger and more fluid modes of organisation. To paraphrase Simon O’Sullivan (2016), a probe-head might be any form of practice that ruptures the dominant regime thus creating something else, something new: a future that wasn't there before.

Holes

Philosopher Joshua Ramey (2018) argues against the neo-liberal appropriation of the future and colonisation of the present, proposing instead a kind of anti-futurity... a refusal of the entire category of the future. If time is somehow in front of us, then, to trouble it what is needed is a non-linear intervention that pierces holes, wormholes in the received notions of what the future should be.

Picklocks

This is where divinatory practices become interesting: not because they offer definitive answers, or a clear-cut decision-making technique, but because they are a process of ongoing inquiry into uncertainty that can accommodate enigmatic, equivocal, or even opposed and conflicting meanings: this ambiguity must be treasured if we want to inhabit the contingency of the world.

It is this act of inquiry that sets possible possibles in motion... cracking new openings to redefine what will have been possible.

The emphasis on divination and conjectural knowledge should not be misinterpreted as a disregard for scientific knowledge, or worse, as a rejection of machinic technologies.

On the contrary, it must be read as a strategic intervention that counteracts the pseudo-science of contemporary data occultism (data mining company Palantir named after a crystal ball in the lord of the rings) and its techno-deterministic allegedly ‘objective’ knowledge-production (Anderson, 2008).

This is our first step into this space, to conclude we would like to leave you with a set of questions that we will be thinking about going forward:

Can machines predict anything other than a coarse conservative version of the past?
How do we contrast the algorithmic manoeuvres that cull the possible by turning it into the probable?
How do we diagram potential, “all those ‘might haves’ or ‘could bes’ implicit in situations”? (Debaise and Stengers, 2017, p. 17).

Anderson, C. (2008) ‘The end of theory: The data deluge makes the scientific method obsolete.’, Backchannel - Wired.

Candy, S. (2010) The futures of everyday life: Politics and the design of experiential scenarios. PhD thesis.

Debaise, D. and Stengers, I. (2017) ‘The Insistence of Possibles. Towards a Speculative Pragmatism’, PARSE Journal. Translated by A. Brewer, 7.

Debaise, D. and Stengers, I. (eds) (2015) Gestes spéculatifs. Dijon: Les Presses du réel.

Deleuze, G. and Guattari, F. (1987) A thousand plateaus: capitalism and schizophrenia. Minneapolis: University of Minnesota Press.

Guattari Félix (2016) Lines of Flight. For another world of possibilities. London: Bloomsbury

Ginzburg, C. (1980) ‘Morelli, Freud and Sherlock Holmes’, History Workshop, 9, pp. 5–36.

O’Sullivan, S. (2016) ‘On the Diagram (and a Practice of Diagrammatics)’, in Schneider, K. and Yasar, B. (eds) Situational Diagram. New York, NY: Dominique Lévy, pp. 13–25.

Parisi, L. (2017) ‘Reprogramming Decisionism’, e-flux, 85.

Pedregosa, F. et al. (2011) ‘Scikit-learn: Machine Learning in Python’, Journal of Machine Learning Research, 12, pp. 2825–2830.

Ramey, J. A. (2016) Politics of divination: Neoliberal endgame and the religion of contingency. London: Rowman & Littlefield International (Reinventing critical theory).

Ramey, J. A. (2018) Personal conversation.

Rouvroy, A. (2013) ‘The end (s) of critique’, in Hildebrandt, M. and De Vries, K. (eds) Privacy, Due Process and the Computational Turn. London: Routledge, pp. 143–67.

Stengers, I. (2005) ‘The Cosmopolitical Proposal’, in Making Things Public. Atmospheres of Democracy.

Whitehead, A. N. (1970) Science and the Modern World. Reissue edition. New York: Simon and Schuster.

The Flower and the Future

2018-09-28T00:00:00Z

I gave this talk on 15^th March 2018 at This Happened London #27: Colossal Dust: Practices of Obsession and Investigation, curated by Marion Lagedamont and Rosie Allen.

My current research is a critical investigation of algorithmic prediction. As a designer, my way of approaching this is through diagrams. The mathematics of prediction are deeply spatial and diagrammatic; they construct multi-dimensional vector spaces from numbers. Predictions are basically operations performed on this space: reducing it, expanding it, cutting through it by drawing lines, smooth curves, planes, hyper-planes, and so on. When you hear the words “Machine Learning” or “Artificial Intelligence” they basically mean very sophisticated diagrams. Hopefully this will become a little bit clearer as I talk through the history of one example.

I am going to start by talking about some iris flowers that were measured in 1936. I will then move to today and how these flowers have been enshrined in prediction mythology. I will finish by zooming out again to show how this is just one example of my bigger obsessive practice of mapping the history of prediction.

Anderson (1936)

In 1936, Edgar Anderson, a geneticist and professor of botany, was trying to figure out the species problem in iris flowers.¹ He was trying to understand the genetic relationships between three sub-species of Iris: Versicolor, Virginica, and Setosa. He suspected that Versicolor was a hybrid of the other two.

He conducted a study, measuring the petals and sepals of the flowers. He summarised the measurements of each flower in ideographs. Here is our entry point into vector space, this is the last time we will see an actual flower in this talk.

Ideographs of 20 plants of Iris versicolor. Average of entire colony (50 for each species) shown in central frame.

”Photograph of three-dimensional model showing the precise geometrical relationship in petal and sepal size and proportions of Iris virginica (left), I. Versicolor (center), and I. setosa (right).” Anderson (1936)

Anderson demonstrated the genetic relationship between the species using a 3D model. Versicolor was indeed between Setosa and Virginica, it was even positioned 2 thirds of the way towards Virginica. In this example, the space’s coordinates are defined by the characteristics (length of petals and sepals) and each species is a vector. Their position in the space tells us something, for example that Versicolor is a hybrid, and that it has twice as many chromosomes from Virginica.

Fisher’s Iris Dataset

The real fun, however, started a year later. R. A. Fisher, a statistician involved with the study, published his own use of the data.² He was less interested in the flowers themselves than with finding statistical ways of classifying the species.

When two or more populations have been measured in several characters, x1, ..., x8, special interest attaches to certain linear functions of the measurements by which the populations are best discriminated. (Fisher, 1936)

In 1936 there was indeed “a special interest in how to best discriminate” between populations mathematically. Classifying people and species was, and still is, a hot topic which I’ll come back to. Other research going on at the time included for example craniometry, with Mr E. S. Martin who was trying to differentiate between sexes using jaw measurements.³

Fisher found a way to separate the species mathematically: functions which split the 4-dimensional space of iris coordinates. The potential for prediction here is obvious, if you can separate species by their measurements and you measure a new specimen, you can tell which species it is likely to belong to. This idea is still very much a the heart of prediction today, which brings us to:

In the decades following Fisher’s paper, the basic idea of classifying a population based on data remained more or less the same. The methods got a lot more sophisticated, thanks in part to the development of computing, and started to be called different names such as “pattern recognition.” Throughout this history, Anderson’s iris data have been used as a test-case for new classification techniques, they became a demonstration tool. This is one example from 1976: ⁴

“Left: Figure 5. A two-dimensional display of two groups of the iris data In the optimal discriminant plane. Right: Figure7. Functional display of multi-dimensional data”

Today, the iris flowers are found in tutorials and demos for most machine learning software.

Decision Trees in the scikit-learn package.

Support vector machines are a more sophisticated way of drawing the boundaries. They don’t have to be linear, i.e. they can be curves—and they can classify points with a lot more dimensions. They are shown here in a Mathworks demo.

And of course neural-networks, which are currently all the rage. This is the ‘getting started’ tutorial from Google’s Tensorflow.

This history is interesting to me for a few reasons. First re-tracing these steps shows that what is currently being sold as ground-breaking ‘Artificial Intelligence’ is actually part of lineages that extend far back in time and bring pieces together from unexpected places such as botany. Some of these mathematical techniques are nearly a century old. They did get amplified by more computing power and exponentially more data, but the basic diagrams remain the same: drawing boundaries between categories.

Second, these lineages are not all pretty, and far from neutral. The original 1936 Fisher paper was published in the Annals of Eugenics. It is fairly easy to imagine how a movement concerned with the genetic optimisation of the human race would be interested in statistical ways to “best discriminate” between “classes”. When you download the paper, you get this warning:

warning on PDF cover

At the time of writing this talk, Google Scholar counts 14,206 papers citing this dataset. This does not include countless computer science classes, blog posts, tutorials, and so on. I would be willing to bet that very few of these carry the warning.

Of course it is too easy and simplistic to draw a direct line between eugenics and whoever is using the iris data today. But at the very least, in line with today’s theme, there is eugenics dust on these flowers. It is completely normalised, overlooked, and gets smuggled into our current systems in all sorts of ways.

Will the Real Iris Data Please Stand Up?⁵

My third reason to find the iris dataset interesting is that they are yet another example that data are neither objective nor static. In 1999 a group of researchers noticed that there were, in fact, a few versions of the iris dataset in circulation, with slightly different numbers. What was supposed to be a stable benchmark had over time, over being copied through the decades, slightly mutated into different species of itself.

These benchmark datasets are actually interesting in their own right. They are an idealised space where everything goes right. The data perfectly fits the prediction problem, and the other way around, which never happens out in the world. Yet they set the conditions of what questions can be asked.

This is another popular example, the MNIST database of handwritten digits, which contains 70,000 images.

MNIST database (LeCun, Cortes and Burges, 1990)

This is possibly the equivalent of Iris, which has only 150 data points, but for research in neural networks. Again, the basic idea is similar: you ‘train’ a model to recognise (classify) images of numbers, then you can predict new numbers from that model. In terms of vector space, these are 28x28 pixel images and the value of each pixel is a dimension, just like the width of the petal was one of 4 dimensions in iris. So MNIST has 784 dimensions, things are getting serious.

The goal here is to try to minimise the error of your predictions. This results in a score which is itself a subject for ranking and classification, as seen in the “high scores” table below. The vector spaces of test datasets like MNIST are also a kind of battleground for prediction performance. As Adrian Mackenzie puts it: ⁶

The field of digits becomes a site of differentiation not only of digits—the machine learners attempt to correctly classify the digits—but of the authority of different machine learning techniques and approaches. They become ways of announcing and delimiting the authority, the knowledge claims, or “truth” associated with the machine.

Ranking of classifiers on the MNIST dataset. yann.lecun.com

To conclude I am going to talk a little bit about how I am investigating these things. The iris dataset is just one example of a much bigger investigation which is definitely obsessive, and tests the limits of my sanity.

True Detective

I am building a visual history of prediction called Counting the Future

I am drawing from the literature on the history of statistics and probability, as well the growing field of critical data studies. My particular focus, as I have mentioned, is on extracting diagrams from this literature. I am putting all of these diagrams in relation with each other on a timeline.

These relations are thematic, for example there are threads on insurance, astronomy, finance, and social physics. They follow citation trails, the evolution of devices such as the Bloomberg terminal, or the transpositions of techniques between fields such as astronomy and finance. Essentially I am drawing a diagram of diagrams, and because it is over 300 hundred years

I am calling this practice:

Each of the nodes on this map is an artefact: a combination of diagrams and quotes for context. Here we have our Irises

There are different ways of viewing this information, for example by browsing through the diagrams as a kind of Tumblr blog layout.

And of course you can follow some rabbit holes yourselves by going through the references.

You can view this work-in-progress at countingthefuture.net

Although I use data visualisation tools for this, I have no ambition to be exhaustive or even remotely objective. I am effectively adding a layer of interpretation to accounts from historians of science and other scholars.

This is only a representation of my own knowledge. The links, as I have said, are drawn according to themes or other factors which I am still figuring out. The size of nodes is relative to the amount of information I have attached to them.

I am entering and collecting all of this information “by hand.” Which makes it less like data, and more like what Johanna Drucker calls capta.⁷ The distinction is that capta are “actively taken” whereas data are considered given. This means this project is forever incomplete, subjective, and possibly wrong.

This is what the epic meta-diagram looks like at the moment. The data are literally stored as a diagram—a graph database. I have chosen this format because it is very flexible and reflects the connected, diagrammatic nature of the things I am trying to investigate. The visible part of the timeline is the green nodes at the top, the rest are the references, images and quotes which extend in a kind of rhizome.

I have built myself an editor so the practice of plotting this map, is actually visual all the way through. Again this echoes Drucker and her “visual forms of knowledge production.”⁸

This is the digital workbench on which I do most of my obsessing. For example here I am trying to figure out who came up with the normal distribution:

Crazy Wall in Person of interest. crazywalls.tumblr.com

The best analogy I have so far for this type of practice is the “crazy wall”—the plot device used in crime films and TV series. This is a board or wall on which the detectives chart their progress with pictures, pins, and string. The main difference is that I never expect to find whodunnit

Instead, I am quite happy to keep constructing this thing as an end in itself, and as a starting point for other projects.

As a kind of archive or publication, I also hope this can be useful to others—that it can provide a new way to access the rich and fascinating literature on the history of prediction, and invite multiple readings and re-readings of the lineages that converge in the current moment.

Anderson, E. (1936) ‘The species problem in Iris’, Annals of the Missouri Botanical Garden, vol. 23, no. 3, pp. 457--483 [Online]. DOI: 10.2307/2394164.↩
Fisher, R. A. (1936) ‘The Use of Multiple Measurements in Taxonomic Problems’, Annals of Eugenics, vol. 7, no. 2, pp. 179--188 [Online]. DOI: 10.1111/j.1469-1809.1936.tb02137.x.↩
cited in Fisher (1936)↩
Chien, Y. T. (1976) ‘Interactive Pattern Recognition: Techniques and Systems’, Computer, vol. 9, no. 5, pp. 11--25.↩
Bezdek, J. C., Keller, J. M., Krishnapuram, R., Kuncheva, L. I. and Pal, N. R. (1999) ‘Will the real iris data please stand up?’, IEEE Transactions On Fuzzy Systems, vol. 7, no. 3 [Online]. Available at https://pdfs.semanticscholar.org/1c27/e59992d483892274fd27ba1d8e19bbfb5d46.pdf.↩
Mackenzie, A. (2017) Machine learners : archaeology of a data practice, Cambridge, Massachusetts, The MIT Press. fn p.142↩
Drucker, J. (2011) ‘Humanities Approaches to Graphical Display’, DHQ Digital Humanities Quarterly, vol. 5, no. 1 [Online]. Available at http://www.digitalhumanities.org/dhq/vol/5/1/000091/000091.html.↩
Drucker, J. (2014) Graphesis: Visual Forms of Knowledge Production., Harvard University Press.↩

SEEING-[:like]->a DIAGRAM

2018-03-27T00:00:00Z

I started 2018 by running a workshop at the MA Graphic Media Design, London College of Communication, as part of a series around the theme of Leakage which included Ruben Pater, FRAUD (Audrey Samson & Francisco Gallardo), and Marwan Kaabour.

image: a subset of the Panama Papers database in Neo4j Browser.

Brief: [PDF: 169KB]
Project Reader on Are.na

My way into the theme of Leakage was through graph databases, specifically Neo4j, the database used by the International Consortium of Investigative Journalists behind the Offshore Leaks such as the Panama and Paradise Papers.

Graph databases are a way of storing data as nodes and edges, they are particularly suited to social network analysis and a range of use cases such as recommendation engines or fraud detection. The Offshore Leaks have become a poster story for this otherwise mundane form of data storage; one where technology enables journalists to uncover the dealings of the rich and powerful.

Fraud detection example using Neo4j and Zoomcharts.

Graphs are often vast and very complex–the Panama Papers database, for example, has 1.040.535 nodes connected by 3.071.104 relationships–yet their structure is easy to grasp compared to relational databases which link data through multiple tables. I was curious about what graphic designers would make of these diagrammatic data structures; especially of Neo4j as it features a visual browser and the Cypher language, an ASCII-art syntax for queries written as small diagrams.

MATCH (o:Officer)-[:BENEFICIARY_OF]->(e:Entity)
return o, e
An example Cypher query

The Panama Papers database in Neo4j Browser.

The workshop was about borrowing this ‘way of seeing’ through graphs and applying it to subject areas chosen by the participants. Graphs have the potential to be both reductive and generative; on the one hand, they see the world only through nodes and links, on the other, they reveal new, hidden, or surprising connections. The brief was about negotiating with these possibilities and looking for creative opportunities. The spectrum of suggested approaches to the brief was delimited by “critical investigation” at one end and “speculative imagination” at the other. In other words, the graphs were not an end in themselves but grounding for critical practice and discursive projects, with the aim to either unpack, comment on, or propose alternatives to, existing ideas, systems, and narratives. This process took place in two phases:

First participants chose an area of interest and set out to map it using a graph. I briefly introduced Neo4j, but teams were free to use other means to construct their graphs—such as collaborative drawing tools or physical pinboards. Since the Neo4j browser doesn’t allow for visual editing without writing code, some teams used APC Jones’ Arrows tool to build their graphs visually. Each group designed ways of finding, accumulating, and systematising data into a graph. This was a labour intensive process which prompted reflections on the act(s) of “actively taking” information—what Johanna Drucker calls capta, as opposed to data which are “assumed to be a ‘given’ able to be recorded and observed.”¹

Slide by Wesley Goatley

The second phase was about designing interventions and proposals. These were drawn, extracted, or otherwise inspired by the graphs, whether directly querying a database or more loosely. The graphs themselves were treated as an outcome in their own right, to be submitted separately. This meant that this phase was explicitly not a graph/data-visualisation exercise. While network visualisations have proven useful, for example in the forensics of twitter bot networks, they often turn into “hairballs” or “spaghetti bowls”². Instead, this part of the workshop was about design which was informed by the graph, but the terms of this translation were up to the participants.

Below are summaries of the different projects and some selected visual outcomes.

Crisps

Team: RuiQing Cao, Huancui Chen, Qianxian Chen, Ruiqi Chen

Graph: Ingredients from a sample of 31 types of crisps. The graph goes from the macro-scale with the industrial groups to which the crisp brands belong, to the micro-scale with the chemical compounds found in each of the products.
Tool: draw.io

Interventions: Packaging designs which highlight the gap between marketing language (“natural”, “hand cooked”) and the actual ingredients found in crisps. These were experiments in foregrounding what is usually hidden in the ‘small print’ of processed foods.

Oil Trust

Team: Chi Kit Chan, Wei Dai, Helen Taranowski, Shuang Zhou

Graph: The locations, companies, and infrastructure involved in 61 oil spills. As the group narrowed their focus to the public relations ‘management’ of these disasters, they attached elements of PR language to some of the events and companies.
Tool: Neo4j

Intervention: Oil Trust, a satirical PR firm specialising in oil spill crisis management. Includes a press-release generator, merchandise store, brochures, and social media accounts.

Art Riot

Team: Xiaoxuan Guo, Rong Tang, Xiaoquing Wang, Hao Zhang

Graph: Relationships between artworks, artists, news events, and broader social themes in Russian activist art, focusing on the Art Riot exhibition (Saatchi Gallery, 2017).
Tools: Arrows, Drawings

Interventions/Proposal: A series of tributes to Russian artists/activists, and a new graph-based way of curating art exhibitions and books through common themes or visual elements. This was explored through a web-app prototype and a publication.

Code of Arms

Team: Aadhya Baranwal, Sui-Ki Law, Clara Wassak, Shengtao Zhuang

Graph: The symbols, colours, and typefaces used in 29 heraldic emblems for EU countries (28 + Scotland), as well as their semantic meanings.
Tools: Arrows, Neo4j

Proposals: A re-interpretation of the semantics of EU national heraldry using emoji, as the basis for a personal emblem generator (see the prototype). This is one of a series of experiments which used the graph as a generator for new emblems; for example for the whole of the EU, or for “unifying” two nations such as the UK and Ireland.

I won’t attempt to draw general conclusions from such a diverse set of subject areas, outcomes, and ways of working. If anything, what runs through these projects is the tension mentioned earlier between the systematic graph, a reductive form of data storage (as they all are), and the generative graph, a map which invites a multiplicity of readings. I was genuinely impressed with how the participants took ownership of this open-ended brief, with the fascinating and challenging subjects they chose to investigate, and with the work they produced.

A series of events based on the Leakage workshops will take place at the Design Museum in June 2018 as part of the Hope to Nope exhibition. The ideas and work presented here will continue there in another, yet to be determined, form.

Huge thanks to all of the participants for taking part and working hard, to Paul Bailey for inviting me to give the workshop, and to Georgina Voss for her feedback on the projects during the final presentations.

Drucker, J. (2011) ‘Humanities Approaches to Graphical Display’, DHQ Digital Humanities Quarterly, 5(1). [link]↩
Bounegru, L., Venturini, T., Gray, J. and Jacomy, M. (2017) ‘Narrating Networks. Exploring the affordances of networks as storytelling devices in journalism’, Digital Journalism. Routledge, 5(6), pp. 699--730. doi: 10.1080/21670811.2016.1186497.↩

Anticipation 2017

2017-11-10T00:00:00Z

I gave the following talk on 10.11.2017 at the Anticipation 2017 conference.

I am a PhD candidate at the Royal College of Art on a scholarship from Microsoft Research Cambridge. My research is a practice-based critical investigation of algorithmic prediction. I come from a critical and speculative design background with the aim to contribute to the broader conversation about the role of data and algorithms in society and culture; also known as critical algorithm studies.

I focus my research on the visual and spatial aspects of prediction as an entry point to bigger social, cultural and political questions. On the surface this starts with data visualisation but it goes all the way down to the multi-dimensional vector spaces on which the predictive operations of machine learning are performed.

I’m going to talk about one of the projects I’m currently working on called The Monistic Almanac. Adrienne Lafrance has suggested that the Farmer’s Almanac was a precursor to the information age. I am extending this parallel to data-science and data visualisation, and applying it in practice by making my own, contemporary version of an almanac.

Almanacs are practical guides to the future in areas such as nautical navigation, farming, finance, and many more. Published for the year ahead, they provide reference points to navigate an uncertain world, and bring a sense of cosmic order to everyday life. They are artefacts from the long history of data-centric predictions. From a design perspective, looking at almanacs helps to start unpacking some of what Barnes and Wilson¹ call “big data’s historical burden”.

Annuaire pour l'an 1875, publié par le Bureau des longitudes² (emoji added)

The notion of monism is a key aspect of this burden, the idea that the same set of laws govern both the natural and social worlds. One example is the way in which the mathematics of astronomy were borrowed in the 19th century and applied to social domains, to predict behaviours like marriage, suicide, or crime, and to legitimise practices like financial speculation³. This is reflected in almanacs as astronomical tables are printed next to interests on loans, all sorts of unit conversions, and so on… Today monism is alive and well, the same computational statistics are applied to virtually every domain across science, business, and society.

My project is about pushing monism to its absurd extreme. The almanac is an interesting site to do this because unlike the grand promises of big data, it doesn’t take itself too seriously. It aims to be "useful with a pleasant degree of humour", with different rationalities coexisting happily, sometimes within the same publication: science, astrology, divination, folk knowledge, remedies, proverbs. Some, like the Old Moore’s Almanac, make predictions with the tone of a tabloid/gossip newspaper.

In practice I am using the tools of data science, machine learning and data visualisation—such as the jupyter notebook, scikit-learn, and D3.js—to construct a series of predictive rationalities; basically a set of machine-learning astrologies.

The first one is called Cosmic Commodity Charts. It uses the positions of the planets of the solar system to predict prices on the commodity markets. It uses a support vector machine, essentially a regression, to ’learn’ the relationship between planet positions and prices using 30 years of historical data. Using future planet positions⁴, it can then produce price predictions.

The second one is the Crisis Proximity Index, an astrology based on the financial crisis of 2008. It takes the 9th of August 2007 as its reference point—the day BNP Paribas froze three of its funds, showing the first cracks in trust in the subprime system. Planet positions and the direction of their movements are interpreted on this basis, anything approaching the positions as they were on that date is considered negative, getting further away is read as a positive sign.

crisis_date = date(2007, 8, 9)

I’ll be expanding on the series in the coming weeks and months, and eventually publish it as an automated website. Sign up for the newsletter at almanac.computer if you want further updates.

This is still in progress as you can see, but some broad themes are already emerging. I’ll run through these briefly to conclude:

Coming back to prediction and my spatial focus, these two modules are performing operations in space; regardless of the planets in actual space. Wether it’s a flattening in the regression of planet distances, or about moving closer or further from a crisis position, these systems are about giving predictive meaning to literal distances. This echoes things like the nearest neighbour algorithm, one of the most obvious examples of the spatial nature of computational predictions.

I’m interested in failure and absurdity as a mode of critical engagement. Karppi and Crawford⁵ write about the ‘Hack Crash’ of 2013 where the twitter account of the Associated Press got hacked and sent out fake tweets about explosions at the White House, which triggered automated reactions on the stock market in the minutes that followed. These were quickly corrected but the researchers note that when systems fail, we get unusual glimpses into how they work. What if we constructed absurd systems that only do this part?

Finally, constructing these rudimentary systems myself is a way to look ‘inside’ at their inner workings, but a lot of it remains abstracted. I am relying on software libraries that do most of the mathematical work for me, which is part of my point. As Ananny and Crawford⁶ point out, the transparency ideal—seeking to understand algorithms by looking inside—has serious limitations. What is more interesting is to look across, in my case at how belief systems get encoded in computational systems.

Barnes, T. J. and Wilson, M. W. (2014) ‘Big Data, social physics, and spatial analysis: The early years’, Big Data & Society, 1(1), pp. 1–14. doi: 10.1177/2053951714535365.↩
Bureau des Longitudes (1875) ‘Annuaire pour l'an ... publié par le Bureau des longitudes’, Available at: http://gallica.bnf.fr/ark:/12148/bpt6k65393410.↩
Ashworth, W. J. (1994) ‘The calculating eye: Baily, Herschel, Babbage and the business of astronomy’, The British Journal for the History of Science, 27(04), pp. 409–441. doi: 10.1017/S0007087400032428. JSTOR ↩
Future planet positions are provided by the Jet Propulsion Laboratory’s DE430 Ephemeris, accessed through the Skyfield python library.↩
Karppi, T. and Crawford, K. (2016) ‘Social Media, Financial Algorithms and the Hack Crash’, Theory, Culture & Society, 33(1), pp. 73–92. doi: 10.1177/0263276415583139.↩
Ananny, M. and Crawford, K. (2016) ‘Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability’, New Media & Society. doi: 10.1177/1461444816676645.↩

Predicting the Future from the Stars

2017-09-18T00:00:00Z

I recorded this interview with Dr. Roberto Trotta at Imperial College London on 18.09.2017. It was aired as part of the Very Very Far Away pop-up radio station at the Victoria & Albert Museum on 23/24.09.2017.

Here are some of the references mentioned:

Ashworth, W. J. (1994) ‘The calculating eye: Baily, Herschel, Babbage and the business of astronomy’, The British Journal for the History of Science, 27(04), pp. 409–441. link

Davenport, T. H. and Patil, D. J. (2012) ‘Data Scientist: The Sexiest Job of the 21st Century’, Harvard Business Review, October. link

"the best thing about being a statistician is that you get to play in everyone's backyard" John Tukey cited in:
O'Neil, C. and Schutt, R. (2013) Doing Data Science. O'Reilly Media, Inc.

Stanley, M. (2013) ‘Where Is That Moon, Anyway? The Problem of Interpreting Historical Solar Eclipse Observations’, in Raw Data Is an Oxymoron. MIT Press, pp. 77–88.

Espenak, F. and Meeus, J. (2006) Five Millennium Canon of Solar Eclipses: -1999 to +3000 (2000 BCE to 3000 CE). NASA. link

Divination is "any ritual and its associated tradition performed in order to ask a more-than-human intelligence for guidance"
Curry, P. (2016) Divination; Perspectives for a New Millennium. Routledge.

Levinovitz, A. J. (2016) How economists rode maths to become our era’s astrologers, Aeon Magazine. Edited by S. Haselby, 4 April. link

James, R. (2015) Cloudy Logic, The New Inquiry, 27 January. link

Counting the Future

2016-09-09T00:00:00Z

Minimum Viable Product, as presented on 01/09/2016 at the Counting by Other Means session (Kember and Taylor, 2016); 4S/EASST 2016 Science and Technology by Other Means Conference in Barcelona, Spain.

10 Apr. 2019: This article was previously hosted on Medium. I am moving it here so it can be self-hosted with the rest of my talks.

Introduction

Last June, an article in Wired magazine was titled:

“This AI learned to predict the future from watching loads of TV”

This story was based on a computer vision study from MIT where machine learning was used to predict handshakes, hugs, kisses and high-fives in american sitcoms with 43% accuracy (Burgess, 2016).

This is but one example of the current explosion in machine learning, which is currently transforming a huge variety of fields, and of the high expectations being set by headlines declaring that long lasting myths such as machine intelligence and precognition are finally becoming reality.

Machine learning, whether it is called ‘data science’ or ‘predictive analytics’, is currently at the centre of an imaginary where computers predict the future. As many of you will know, it promises everything from great profits, with Eric Siegel describing predictive models as ‘golden eggs’ — which don’t even have to be very good, as predictions just marginally better than random guesses will already pay off (Siegel, 2015)— to the obsolescence of the scientific method (Anderson, 2008).

‘Anticipatory Design’ proposes a user interface for the predictions of machine learning, or more accurately it proposes to remove as much of the interface as possible by making decisions for the user to erase the mountain of tedious choices out of modern daily life (Shapiro, 2015).

This project proposes to use design practice to critically examine computational prediction; to offer a counterpoint to anticipatory design by going back in time and looking at history. However groundbreaking the current moment is for computational prediction, it is the latest chapter in a long history of attempts to predict the future from data. Looking into this history might reveal some of the recurring narratives and aesthetics that define the systems that we find ourselves entangled in today. To do this I am building a visual history of prediction technologies; a web based index of attempts at predicting the future from data.

This paper is about the first ‘minimal viable product’ for this research tool and some of the things I am interested in finding out with it. It sets out some methods for a design-oriented investigation of computational prediction–ways to plot knowledge from the literature on the history of probability and statistics, borrowing methods from the digital humanities — and proposes some early findings, or hunches, about the narratives and aesthetics of predictions.

I. Methods

Rather than attempting to re-write an exhaustive history of prediction, which others have done very well (Porter, 1986; Gigerenzer et al., 1990; Daston, 1995), this project aims to contribute new ways of accessing this history by approaching it from a design perspective. Both the questions being asked, such as ‘how does prediction manifest itself aesthetically?’, and the methods used to address them are rooted in design practice. The ways in which we extract or ‘mine’ predictions from data is here both a subject to be critically examined and the means of an investigation to find ‘other means’ of ‘counting’ this history.

Artefacts as data points
As a designer I aim to uncover the characteristics and aesthetics of prediction technologies through the study of artefacts. Presented here is a collection of objects and stories which embody specific attempts at predicting the future from data. They provide tangible glimpses into the long and intricate history of statistics, using “objects” as an entry point to “values” and “rules” (Gigerenzer et al., 1990). As objects of design, the artefacts in Counting the Future are selected for the connections they provide to the broader designs (as in motivations, politics and imaginaries) that they are shaped by.

Data Stories of Data
To experience the complications of ‘mining’ data for knowledge and patterns first hand, I am using a popular data-visualisation library, D3.js to draw chronological and narrative links between the artefacts. The visual language of data and statistics is used to tell the story of how they came to play such an important role in society.

This design process aims to surface some of the complications of computational prediction which are reflected in the practice itself. The problem of how to represent the complex web of technologies, politics and beliefs in which computational predictions exist raises issues familiar to the digital humanities, as described by Burdick et al.: “computation depends on disambiguation at every level, from encoding to the structuring of information. Explicit step-by-step procedures form the basis of computational activity. However, ambiguity and implicit assumptions are crucial to the humanities” (Burdick et al., 2012, p. 17).

The Timeline as ‘Crazy Wall’

To address these challenges, I am working on the beginning of a visual language for the entangled trajectories of prediction. The artefacts are categorised into four narrative themes (see below) but can also be be assigned an additional theme to reflect analogies and influences gathered from the literature. When that is the case the two visual strands become visually connected creating a ‘knot’ in an otherwise linear and parallel timeline. This very rudimentary system is a first step towards presenting the history of prediction as complex and entangled rather than as a linear progression of technological capabilities.

Visually, the result resembles the ‘crazy walls’ used as narratives devices in movies and TV series where a mystery or investigation needs to be solved (Benson, 2015). These boards or walls are used by protagonists, usually detectives, to gather evidence and draw links between suspects as the plot progresses. This also reflects my process as a designer venturing in to the world of technology studies and the history of science, with the difference that I never expect to find whodunnit; the crazy wall here is an end in itself.

II. Narratives and Aesthetics

This project aims to contribute more than a catalog of artefacts as proofs that attempts to predict the future from data are numerous and diverse. I aim to draw on these examples to reveal some of the characteristics and aesthetics of prediction.

1) Themes

I have identified four themes as an initial classification. They are both physical scales and narrative arcs which run through the history of prediction: the cosmos, the market, society and the body. There is no time to go into detail here so I will only give a quick overview of them:

#HEAVENS Cosmic truths and atmospheric forecasts
This theme starts at the origins of scientific prediction with the plotting of planetary orbits from observations, for example the predictions that comet Haley would return in 1759 (Broughton, 1985). It extends through the history of weather forecasting, a poster story of prediction, to today’s climate and ecosystem models (Edwards, 2010).

#GAMBLES Market speculations: bets, bluffs, corners.
The second theme looks at tensions between gambling and economics, as futures became an object of trade. It starts with Chicago’s grain futures market at the end of the 19th century (Cronon, 1991), and continues through the explosion in derivatives trading (D. Mackenzie, 2008) and the high frequency trading we know today. ‘Bets’ about what a customer might want, such as targeted advertising and algorithmic recommendations (Seaver, 2012), are also filed under this category.

#MASSES Social order and public opinion
This theme is about attempts to understand: 1) what society is: a galaxy, as suggested by Quetelet’s Average Man? (Gigerenzer et al., 1990, p. 40) or an organism, following Galton’s analogy between gemmules (an early name for genes) and voters in a constituency? (p.55) and 2) what it wants, for example through political polls and the notion of ‘public opinion’ (Lepore, 2015); in the hope to iron out outliers and imperfections such as criminals (Gibson, 2002; Harcourt, 2006).

#FATES Cell fates and individual destinies
Finally, the last theme is about biological notions of predestination. It starts with the 19th century life insurance industry (Bouk, 2015) and continues with genetics and cracking the ‘code of life’, attempts to unlock the secrets of the brain, predicting romantic attraction, and so on.

General Laws of Motion
The themes work as individual narratives but overlaps and intersections surface when they are taken as a series. I am highlighting these with the rudimentary visual system shown above.

They reveal that the history of prediction is one of analogies and translations where theories and formulas developed to solve a specific problem in one field, such as predicting the position of a star in astronomy, were then used in a completely different context such as predicting crime in society; as in Adolphe Quetelet’s ‘Average Man’: “The average man was invented as a tool of social physics, and was designed to facilitate the recognition of laws analogous to those of celestial mechanics in the domain of society.” (Gigerenzer et al., 1990, p. 40).

The fact that identical mathematical methods are applied to a wide variety of different domains stands out as key characteristic of computational prediction. Coming back to machine learning, Mackenzie talks about “a generalization of prediction […] woven into the fabric of everyday life.” (A. Mackenzie, 2015, p. 430). On top of the specific narratives in the themes above, the overarching idea is that the same general laws of motion govern the cosmos, markets, society and bodies, and that these can be accessed through universal and interchangeable mathematics.

2) The Aesthetics of Accuracy

Coming back to my design focus, I’m interested in how these universal mathematics, or ‘codes’, of prediction are represented. They are numbers first and reflect “the fascination with the numerical and the longing for certainty that the numerical symbolizes, preconditions for the remarkable success of the mathematics of uncertainty.” (Gigerenzer et al., 1990, p. 237), but they are also inherently visual.

Visual Codes
The rise of probability and statistics in the second half of the 19th century coincides with a “golden age of statistical graphics” (Friendly, 2008), an explosion in the types and numbers of charts, graphs and plots. These were not just illustrations to communicate numbers but a key element in the production of predictions. From a graphic design perspective, the history of prediction can be summarised as a long quest to draw smooth curves through noisy data points and to reveal patterns through plotting.

The Legitimacy of Codes
Aside from the graphical aspects of mathematics, the aesthetics of accuracy also legitimise predictions and its codes. Just like economics borrowed legitimacy from astronomy in the 19th century by using its rigorous mathematics (Ashworth, 1994), the codes of prediction lend the appearance of the general laws of motion to political projects; for example this classification of risks for insurance from 1903 includes skin colour, profession, place of birth or residence and profession as risks:

Today’s predictions continue to rely on this legitimacy, implying the general laws of motion through processes of codification. For example finding a romantic match on OkCupid or determining wether a suspected criminal is likely to offend again using the COMPAS score both rely on lengthy questionnaires.

On the face of it, agency is transferred from user to machine through a simple sequence of input > submit > black box > score. The aesthetic of interaction is more of a mutual production of the count where the agency oscillates back and forth between user and machine. This process gradually suspends disbelief so that when the score comes out it is the product of both the specific user and the general laws of motion.

The Befuddled Many
We have seen some of the ways in which the aesthetics of accuracy communicate predictions, but they are as much defined by their opacity than by their appearance. Sandvig writes about the ambiguity of the term ‘algorithm’ and the ways in which ‘the sort’ is represented and absented “cloaked by the legal protections of intellectual property and corporate personhood” and about the need for a “countervisuality” of algorithms to address this opacity, and open up the count for scrutiny (Sandvig, 2014). Sandvig, as well as Pasquale’s Black Box Society, both respond to the sense of befuddlement about what comes to count in the computation of prediction.

This is interesting because befuddlement and uncertainty is exactly what statistics and predictions set out to reduce in the first place:

“Only a small elite of hommes éclairés could reason accurately enough by unaided intuition; the calculus of probabilities sought to codify these intuitions (which the probabilists believed to be actually subconscious calculations) for use by hoi polloi not so well endowed by nature. This mathematical model of good sense could be compared to spectacles. By applying the same optical principles responsible for normal eyesight it was possible to extend vision artificially; similarly, the calculus of probabilities formalized the good sense that came naturally to the fortunate few to help out the befuddled many. ”(Gigerenzer et al. 1990 p.16)

Although this quote refers to the enlightenment and what Daston calls ‘Classical Probabilities’ (1995), it resonates with some of the aesthetics of computational prediction today, such as Anticipatory Design’s mission to help out users of technology crawling under “decision fatigue” by making decisions for them (Shapiro op. cit.).

It poses a question that could be asked about today’s systems of prediction: who are the fortunate few and the befuddled many? In the case of Google Now (pictured), there seems to be a simple answer that the fortunate few are ‘affluent white males’ in silicon valley (Clark 2016) helping out the befuddled rest of us. In other cases such as the COMPAS score, befuddlement seems distributed in more complex ways, since the score is provided by a third party (Northpointe Inc, 2016) the suspected criminal and the judge are both part of the befuddled many; in different ways and of course with very different consequences. As I attempt to map out this space I definitely count myself as part of the befuddled many.

Conclusion

I hope to have given you an overview of some of the hunches I will work to substantiate in the coming months and of the ways design practice can be used to approach prediction and its complications from another angle. There are still many dots to be connected on the timeline which as of now counts 18 artefacts. Work in the coming months will involve adding more artefacts, identifying more knots between the themes, and developing the visual language. One feature in development is the possibility to highlight more specific narratives and transitions, rather than link all artefacts by default.

As part of my practice-based phd project as a whole, Counting the Future is Phase 1 of 3. The next phase will be to experiment and manipulate these aesthetics of accuracy ‘hands on’. One format for doing this might be the iPython notebook which allows for executable code to be mixed with text and images. Peter Norvig uses this format to reimplement some of the classical examples of probability in Python (Norvig, 2016).

I wonder how re-enacting historical theories such as Quetelet’s Average Man with today’s tools such as iPython would bring out differences but also similarities with today’s imaginaries of prediction.

I welcome any questions, comments or feedback, especially as this my first time presenting this work. So feel free to get in touch or to view the timeline here: countingthefuture.davidbenque.com .

(as the site is likely to change, code a the time of writing from which screenshots in this document are taken is accessible here: https://github.com/davidbenque/Counting-the-Future/releases/tag/4S%2FEASST2016)

This project is part of my PhD research in Information Experience Design at the Royal College of Art in London. This work is supported by Microsoft Research Cambridge through its PhD Scholarship Programme.

References

Anderson, C. (2008) The end of theory: the data deluge makes the scientific method obsolete., Wired. Available at: http://www.wired.com/2008/06/pb-theory/ (Accessed: 2 December 2015).

Angwin, J., Mattu, S., Larson, J. and Kirchner, L. (2016) Machine Bias, Propublica, 23 May. Available at: https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing (Accessed: 12 July 2016).

Ashworth, W. J. (1994) ‘The calculating eye: Baily, Herschel, Babbage and the business of astronomy’, The British Journal for the History of Science, 27(04), pp. 409–441. doi: 10.1017/S0007087400032428.

Benson, R. (2015) Decoding The Detective’s “Crazy Wall”, Esquire.co.uk, 23 January. Available at: http://www.esquire.co.uk/culture/film/news/a7703/detective-show-crazy-walls/ (Accessed: 5 August 2016).

Bouk, D. (2015) How Our Days Became Numbered. University of Chicago Press.

Broughton, P. (1985) ‘The first predicted return of comet Halley’, Journal for the History of Astronomy. Available at: http://articles.adsabs.harvard.edu/cgi-bin/nph-iarticle_query?1985JHA....16..123B&data_type=PDF_HIGH&whole_paper=YES&type=PRINTER&filetype=.pdf.

Burdick, A., Drucker, J., Lunefeld, P., Presner, T. and Schnapp, J. (2012) Digital Humanities. Cambridge MA, London England: MIT Press.

Burgess, M. (2016) This AI learned to predict the future by watching loads of TV, Wired, 22 June. Available at: http://www.wired.co.uk/article/algorithm-watch-tv-predict-future (Accessed: 26 August 2016).

Cronon, W. (1991) ‘Trading the Future: Grain’, in Nature’s Metropolis.

Daston, L. (1995) Classical Probability in the Enlightenment. Princeton University Press.

Edwards, P. N. (2010) A Vast Machine. MIT Press.

Friendly, M. (2008) ‘The Golden Age of Statistical Graphics’, Statistical Science, 23(4), pp. 502–535. doi: 10.1214/08-STS268.

Gibson, M. (2002) Born to crime. Praeger.

Gigerenzer, G., Porter, T., Swijtink, Z., Daston, L., Beatty, J. and Kruger, L. (1990) The Empire of Chance. Cambridge University Press.

Harcourt, B. E. (2006) Against Prediction: Profiling, Policing, and Punishing in an Actuarial Age. The University of Chicago Press.

Kember, S. and Taylor, A. (2016) 4S Preview: Counting by Other Means, 4sonline.org, 4 August. Available at: http://www.4sonline.org/blog/post/4s_preview_counting_by_other_means (Accessed: 10 August 2016).

Lepore, J. (2015) ‘Are Polls Ruining Democracy?’, The New Yorker, 16 November. Available at: http://www.newyorker.com/magazine/2015/11/16/politics-and-the-new-machine.

Mackenzie, A. (2015) ‘The production of prediction: What does machine learning want?’, European Journal of Cultural Studies, 18(4–5), pp. 429–445. doi: 10.1177/1367549415577384.

Mackenzie, D. (2008) An Engine, Not a Camera. MIT Press.

Northpointe Inc (2016) COMPAS Core. Available at: http://www.northpointeinc.com/products/core (Accessed: 12 September 2016).

Norvig, P. (2016) A Concrete Introduction to Probability (using Python), Jupyter nbviewer, 12 February. Available at: http://nbviewer.jupyter.org/url/norvig.com/ipython/Probability.ipynb (Accessed: 27 April 2016).

Porter, T. M. (1986) The Rise of Statistical Thinking, 1820–1900. Princeton University Press.

Sandvig, C. (2014) ‘Seeing the Sort: The Aesthetic and Industrial Defense of “The Algorithm”’, Media-N. Available at: http://median.newmediacaucus.org/art-infrastructures-information/seeing-the-sort-the-aesthetic-and-industrial-defense-of-the-algorithm/.

Seaver, N. (2012) ‘Algorithmic recommendations and synaptic functions’, Limn, (2).

Shapiro, A. (2015) The Next Big Thing In Design? Less Choice, fastcodesign.com, 15 April. Available at: https://www.fastcodesign.com/3045039/the-next-big-thing-in-design-fewer-choices (Accessed: 28 August 2016).

Siegel, E. (2015) How Predictive Analytics Delivers on the Promise of Big Data, 23 January. Available at: https://www.youtube.com/watch?v=CtNCuL39AWo (Accessed: 10 March 2016).

Vondrick, C., Pirsiavash, H. and Torralba, A. (2015) ‘Anticipating the future by watching unlabeled video’, arXiv.org.