# 21/01/2022 - The Earth May Not Be Flat, But Our World Increasingly Is

It's funny how the colours of the real world only seem really real when you viddy them on the screen - Alex, A Clockwork Orange

# The Flattening of Reality

# Biological Dimensionality Reduction

In Kant’s Critique of Pure Reason, he distinguishes between noumena (things in themselves) and phenomena (things as they appear), arguing that subjects can never truly have access to objects, only their appearances. The world appears to us as representations given by our senses and constructed in our minds, but these representations are incomplete and, by definition, subjective.

A simple example of this is colour. We perceive colour as a continuous property of physical objects but this is just the way our minds represent variations in the electromagnetic radiation reaching our eyes. My red may not look the same as your red, and we know for a fact that it looks different to my dog’s red. In Kant’s estimation, even time and space are not real things that exist independently of us, rather they are subjective structures that enable perception. While this subject-object disconnect precludes any possibility of objective reality or truth, it is only through our subjective faculties that understanding is even possible in the first place. The world is a firehose of noisy information that our mind filters, parses and makes sense of.


Image by Nathan Saad.

To further illustrate this, let us consider the experience of standing on Bondi Beach in Sydney. I am being bombarded with sense-data - the heat of the sand on the soles of my feet, the sounds of seagulls, the vision of waves crashing against the shore - but this is still only a representation of the real world. There is electromagnetic radiation outside of the human visual spectrum which I can not perceive and sounds my ears are unable to pick up. This is to say nothing of the theoretically perceivable information that I simply miss - for example, the man shaking sand off his towel behind me or the conversation taking place 100 metres to my left. Experiencing the world is always an exercise in information loss.

This biological procedure of filtering and prioritising information to facilitate processing has been replicated and extended through human culture but it has been super-charged in the colonisation of digital spaces. As our lives have moved increasingly online, in-person rituals have been digitised, physical artefacts vectorised and human activities automated. This cyber-emigration has necessitated informational downsampling and, as with Kantian cognition, we have remodelled the world to fit our particular computational framework. Unlike its organic predecessor, this version of pruning was not shaped by evolutionary forces but rather by technocrats, capital and CPUs.

# Mathematical Dimensionality Reduction

There is an analogous procedure in mathematics called dimensionality reduction where high-dimensional data is transformed into a lower-dimensional space. These algorithms aim to reduce computational costs while retaining most of the important information in a dataset. To help build some intuition around this process I have generated the example below. In the top plot, we can see 21 observations represented in 3-dimensional space. Using a technique called Principal Component Analysis (a linear dimensionality reduction method) the dataset is mapped to 2-dimensional and, finally, 1-dimensional space. We can see that while information has been lost, the overall structure is retained.


Dummy example of Dimensionality Reduction using Principal Component Analysis.

A related concept is intrinsic dimension, which is the minimum number of latent variables required to get a good approximation of a dataset. We can roughly think of this as a measure of content or information (I have written before about Claude Shannon's work on Information Entropy) and recent research by Bailey, Houle & Ma (opens new window) has in fact shown that there is an asymptotic relationship between the intrinsic dimensionality and entropy of a dataset. While dimensionality is only technically applicable to matrix datasets, we can still conceptualise the intrinsic dimension of non-tabular representations even in instances where we can’t quantify it.

To get a better sense of how this relates to real-world representations, let’s return to the Bondi Beach example. Perhaps I had a friend who was unable to join me on the beach that day but who wants to know what it was like. I could send them a long email describing the day and they should, assuming my writing is evocative enough, be able to imagine my experience. Something has been communicated, but the email has a lower intrinsic dimension than my information-rich experience. A digital photograph would also be a dimensionality-reduced representation of my day at the beach. In this example, we can actually quantify the number of dimensions: the number of pixels in the photograph multiplied by 3 (corresponding to red, green and blue light intensities). We could even reduce the day down to a single dimension if we wanted to, such as the average air temperature or the number of patrons on the beach. Each of these is a low-dimensional representation of my experience but in every case, information has been lost.

# Ghosts of Translation

If information is lost, what does this residual look like? By definition, it is a lack. It is the subtle difference we notice between a Zoom call and a face-to-face conversation. It is the difference between listening to a vinyl record while perusing giant tactile cover art and hitting play on Spotify. These are the ghosts of translation. Sometimes components of this lack can be expressed - for example, we know that sarcasm tends to be easier to identify in the flesh than in text-based communications because in the latter case we lose the dimensions of vocal tone and facial expression - but often this residual is hard to capture in words, and this is for good reason.

As our lives were progressively commodified and flattened by the market, we replaced material objects and rituals with digital products and those products were designed to meet the needs that we could name. But the less definable needs - the ones we couldn’t name - have been abandoned. For instance, the sense of community humans traditionally felt in churches and sports clubs has largely been replaced by group chats and feed scrolling. We all know implicitly that these generate a cheap imitation of the relatedness we used to feel, but we do not have the vocabulary for that hole and so it was never catered for. We took an ontological cleaver to the world and expunged all that could not be articulated, digitised and monetised.


These developments have plainly not been without their benefits, but things especially degenerated when we started applying this extreme dimensionality reduction process to ourselves. We metricated our performances, quantified our existence and rubricated our behaviour. All facets of humanity that could not be liquified and poured into a database table were jettisoned until every student was an ATAR, every employee a list of KPIs, every economy with millions of agents a GDP result.

This shift has been exacerbated by computational advances but it did not originate here. For example, for decades and across many countries, governments have implemented schemes to distil educational performance. In Australia, students receive an Australian Tertiary Admission Rank (or ATAR) and this single number determines which university courses and jobs are available to them once they leave school. A student's level of understanding, competence and work ethic across multiple subjects (each of which has sub-components) over 12 years of study has high intrinsic dimensionality, yet we reduce this to a representation with 1 dimension. Of course, the ATAR can be a useful proxy. A prospective employer could never take the time to wade through each candidate's reams of exam papers and interview all their teachers, but importantly the employer (in general) does not lose sight of the fact that the ATAR is only a stand-in for the thing it represents. This is why almost all job searches contain an interview component, where the employer can obtain higher-dimensional data about a candidate. But increasingly, and across more domains, we have begun to believe and act as if the representation is more real than the object.

Sign Up To My Mailing List

Hey you are reading this thing! If you are enjoying the article then you should definitely subscribe below. I may be a bot but I will never spam you.

* indicates required

# Upshots of Downsampling

When complex entities are replaced by shallow abstractions, we run the risk of optimising for the wrong thing. Tony Blair and his phalanx of technocrats were an early example of how compulsive metricating can malfunction. In the late 90s and early 2000s, New Labour began allocating funding to public programmes across the UK based on benchmarks concocted by statisticians and policy wonks, but the results were appalling. One of many examples was a financial incentive for doctors who saw patients within 48 hours. At the time the National Health Service was struggling to meet demand, but because of this target, physicians began refusing to make appointments more than 48 hours in advance, lest they lose their bonus. If a patient wanted to make a booking for next Wednesday, they would need to wait until Monday before calling up and hope there was availability. The propensity to meet patient needs - a concept with high intrinsic dimensionality - had been reduced to a single number (percentage of patients seen within 48 hours) and as far as the bureaucrats were concerned, this representation was all that existed.

We see a related effect in the desolate entertainment available to us. Today's large film studios robotically assemble movies guided by focus groups, algorithms and historical data in an effort to de-risk. But while Marvel films successfully appeal to the largest possible number of people (the metric studios are optimising for), they fail to emotionally affect anyone. They entertain in the coldest most technical sense, much like a rat in a cage continuously hitting the orgasm button, but there is a vague yet profound emptiness to all of these products. Again we find ourselves optimising for the limited targets that are easily expressed.

The downsampling of our existence is also dangerous because it generates an illusion of objectivity. Unlike algorithmic dimensionality reduction, our modern versions are often incredibly subjective and can consist of simply picking and choosing numbers that suit our purposes. In the past, when your manager carried out a performance review at work there was no pretence that the outcome was impartial. This was one person's opinion of your productivity and you were well within your rights to disagree with it. But the ubiquitous adoption of KPIs and OKRs has shifted the power-dynamic so that now the performance review results are irrefutable. How can you argue with impartial, unfeeling statistics?

On the topic of power, the flattening of representations has been coupled with a broadening in scope. Our data has shrunk in depth but increased in breadth through the capture and relinquishment of heart rate readings, advertising click patterns, GPS locations, Netflix preferences, step counts, instant messages, holiday photos etc. Governments and corporations today have more knowledge of their citizens/customers than ever before and this affords them an indeterminate, yet palpable power over them. For Foucault, power and knowledge were so inextricably linked that he combined them into a single term - power-knowledge (pouvoir-savoir). This power-knowledge should not be thought of as the ability to control and surveil specific individuals (though in a post-Snowden world we know that this is happening on some level) but rather it is a diffuse capacity to bend the world to their cybernetic designs. While politicians, marketers and others who seek to mislead us, have historically had to rely on rhetoric, now they simply choose the mode of dimensionality reduction that best suits their purposes and pretend that the inference they are drawing is empirically based.

It should be noted that these developments were not entirely sinister in origin. In part, this phenomenon was a quixotic reaction to very real societal flaws and was born of an honourable desire for objectivity in formal settings. For example, by adding stringent criteria to hiring processes, we hoped to excise nepotism and limit the role of unconscious biases (be them sexist, racist etc.). Regrettably, this shift has largely been superficial, with our data-driven utopia the product of a change in methodology rather than values.

# Welcome to the Screenshot of the Desert of the Real

In Charlie Kaufman's Synecdoche, New York, Philip Seymour Hoffman plays a theatre director named Caden Cotard who builds a living replica of New York City within an ever-expanding warehouse. He creates meticulous sets and casts doppelgangers to embody real people until the territories of reality and representation are indistinguishable. While Cotard ultimately chooses to live out his days in his fictional enclave, the situation for us is arguably much worse. In the film at least each stand-in was derived from a real person, place or thing, but our modern experiential compression is more akin to Baudrillardian hyperreality where signifiers no longer require an origin.

The blurring of reality with representation was intensified by technological advances but it did not begin here. In the late 20th Century, Baudrillard's conception of sign-value (and its usurping of use-value) helped decode the unrealness of post-Bretton Woods global economies, celebrity culture in which fame can only be explained recursively, rampant consumerism and wars that exist solely on screens. Our modern condition is simply the natural evolution of this process.

Simplified representations are not intrinsically bad. They are how we understand the world, even at the level of raw sensual observation. Whether organic, algorithmic or manual, dimensionality reduction is an indispensable weapon but we must not ignore the downsides of wielding it. Tech titans, bureaucrats and hucksters have taken a steamroller to reality and if we want to arrest its momentum we should acknowledge the ghosts of translation even when we can not describe them and remain alert to charlatans who use data and statistics to deceive.

If you enjoyed reading this post why not sign up to my MAILING LIST? (opens new window) You'll love it.