Grapple with reality to break out of these shackles

— Brockhampton

While the mechanical phonograph cylinder was invented in 1877, it took almost an entire century before Edison's device was released from His Master's Voice. Sound Engineers and musicians were so endeared with the pipe dream of High Fidelity, that they completely overlooked the creative possibilities that the grammophone also offered.

Once Pandorra was out off the box — and no, it weren't John, Paul, George & Ringo that released her — music would never be the same again. Sound engineers projected the recording arts beyond the spatial and temporal boundaries of reality. Or, to put it in the lingo of Sun Ra, Music was no longer earth-bound; space became the place.

It is high time to free the discourse of Artificial Intelligence from its shackles of fidelity too. Technically, the field has been ready to do so for quite a while. Unsupervised Machine Learning is the branch of machine learning that does not need a human baseline to improve its findings, but rather recognizes patterns that are beyond the scope of human cognition and categorization.

Stubborn creatures as they are, however, humans are often reluctant to accept that their intelligence is surpassed by their own creation. Instead, they rather hold on to their primitive and biased opinions and observations. This post argues that this should change.

Toe the Line

Yo, Dre!
Give me a funky-ass bassline!
Right here!

— Eazy-E

Supervised learning is the most commonly used branch of machine learning. The supervision in its name, designates the practice of developing a human baseline. This baseline is operationalized in the feedback loop between the training data and loss function. For those not familiar with this process and these terms, I'll briefly explain it.

Like most unborns, a machine learning model knows nothing about the world at its inception. To improve, it needs to be taught right from wrong, or at least better from worse. And who better to teach her than it's creator? Right?

In order to train a model, programmers feed a model a set of problems with known answers (labels). These anwers, however, are hidden to the model. At the end of each learning cycle (epoch) the anwers that the algorithm comes up with are compared to the actual answers, and the difference in accuracy (or a similar metric) is measured.

The established baseline is the alpha and omega of supervised learning. The computer will never come up with answers that it has not been taught. But what happens if we were to take the human baseline out of the loop? This process is called: unsupervised learning. In the remainder, I'll only focus clustering — a specific application of unsupervised learning — but the underlying principles hold for other techniques too.

Low-Fidelity

Take the bass line out

— Jay-Z

Unsupervised learning is amoral: there is no right and wrong. In fact, there is not even a clearly defined set of problems or questions; just data points. These data points are converted to vectors, which for the course of this explanation, are best thought of as coordinates. A unsupervised learning algorithm simply measures the distance between the coordinates before it groups them in clusters on the basis of their proximity to each other.

The classic use-case for this form of unsupervised learning is in marketing. Where marketeers used to work with user personae, imaginary models of how their different customers look like and behave. Machine learning turned this process upside down. Instead of defining the models upfront, it discovers clusters of customers that are present in the data. While this was bad news for the story tellers amongst the marketeers, this cluster based targeting turned out to be much more effective… although occasional happy accidents do occur, of course.

In marketing, the benefits of unsupervised learnings are easy to sell. Simply because the human baseline is of less importance than another measurable goal: the bottom line. In other domains, we are often less inclined to believe that computers can come up with better conclusions than us.

Just imagine the following:

• The music of Johan Sebastian Bach has more similarities to that of Justin Bieber than to that of Wolfgang Amadeus Mozart

• The most effective way to solve the climate crisis is the global redistribution of wealth

• President Trump only tweets proven facts

• The earth is in fact flat

These findings are, of course, made up — although I believe at least one of them to be true. They are here to illustrate the following point: are we willing to accept the superiority of an algorithm when its conclusions are counter to our own observations, opinions or beliefs? In other words, are we capable of dropping the human baseline? Lower our voice to make room for better ones?

We should!

Machine Learning does not need to repeat and hollow out human intelligence. It should be used to explore new and better ways to be intelligent. The only thing it needs to unleash its full potential is our permission to drop the baseline.

Photo by Mick Haupt on Unsplash