Instructions for Humans - Day 5

Quiet day in the gallery today, which was handy as I have a sore throat. So much talking last week! I knuckled down and worked on one of the biggest challenges: explaining Machine Learning in the simplest terms possible.

As well as it being a good thing to do, this is an important part of the work. I'm really interested in the practice of acheiving clarity through explaining. Through teaching photography to absolute beginners I've developed a nuanced and clear undertanding of the fundamentals of cameras and composition, which has paid huge dividends in my other work.

I'm seeing this as working an idea or concept like one might work a piece of wood, carving and smoothing it to perfection over time. Teaching is not simply about communicating base facts - it's about finding a vehicle for those facts that seeds them and alows them to grow. I'm really interested in how that process of refinement can be considered an art practice, or at least might lend itself to one. It's something I want to explore over the next few months anyway.

So, here's the first release of Pete Explains Machine Learning. (It's written to appear on slides on a monitor in the gallery which switch every 15 seconds.)

What is Machine Learning?

An all-too-brief and incomplete explainer by Pete Ashton

Artificial Neural Networks

Some early Artificial Intelligence researchers sought to mimic the neurons firing in the human brain with computer systems called Artificial Neural Networks.

Like many experiments in AI, this was a dead-end, but the technology did turn out to have other uses.

As companies like Google found themselves needing to process vast quantities of data on their immense computer farms, they looked to the neural network as a possible tool.

Turns out that while they don't produce consciousness, they're very good at churning through data and finding patterns.

Dumb Maths Done Lots

An artificial neural network is a lot like an ant colony. Individually an ant is an idiot. But collectively they can achieve amazing things.

Each neuron in the network is performing a very simple operation. But there can be millions of neurons and performing billions of calculations in trillions of permutations. This produces a massively complex system.

Patterns in the Statistics

A crude, but accurate, description of Machine Learning technology is statistical probabilities, but on a scale never seen before. Neural Networks find patterns in the data, and then use those patterns to find more patterns.

This can be information we didn't realise was there and makes ML a powerful tool for connecting identities and behaviour in everyday data.

A CCTV system might not see your face but could identify you from your gait as you walk. A website could see patterns in how you move the mouse or the rhythm of your typing.

Digital Assistants with Secret Agendas

Machine Learning systems power most of the internet services we interact with. Amazon's recommendation system churns through product data to match it with your buying patterns. Siri learns to understand the quirks of your voice over time. Gmail studies trillions of messages to learn what spam is. Spotify explores millions of playlists and listening habits to delivery their uncanny personalised weekly playlist.

This is useful, and it's hard to imagine an world without it, but the complexity can hide the workings. Why does Google Maps choose to show you those particular businesses? What isn't it showing you? Why not?

Is more better? Or just more?

A common refrain from Machine Learning researchers is they just need more data to iron out the weirdness and biases in their systems. But how much more data? And what sort of data?

Information is not neutral in itself. The process of capturing, processing, formatting and storing the information can bring in all manner of biases.

No matter how broad the dataset, the system is closed. It is only able to contextualise within the information it has. So it cannot identify bias it hasn't been told about.

It'll probably be alright in the end

Machine Learning is a new field with huge potential to change the world. This makes it exciting and terrifying, particularly when the world is a entering a scary period.

Will this technology be used to bring about a Star Trek utopia or a Philip K Dick hellscape? Probably a bit of both, depending on your reserves of power and privilege.

Where Machine Learning definitely has value is in holding a mirror up to humanity. We can learn so much about ourselves, if we have the guts to look.

Thanks to Jez Higgins (of Hummus zine fame) for casting an expert's eye over this.

Up next Instructions for Humans - Day 4 The impossibility of comprehending the immensity of objective reality, and other games. Instructions for Humans interlude: Project Oversight Turns out Marvel's Captain America movie is the best illustration of data abuse in popular culture.
Latest posts Starmer has effectively introduced a Twitter policy We desperately need social media literacy or the fascists will win History is constantly being “erased”, and that’s OK. Last week a jigsaw saved my life Lost in Bluster Insomnia Journal 07 May 20 Stirchley Safari Whales on the Hudson Closing the Covid browser tabs Solving (a bit of) the Coronavirus from home Stopping A proposal to support independent businesses that cannot survive social distancing A modest proposal for slowing down cars Cross City Walks - the movie Good composition is ideologically fraught Short Reviews of Films Understanding Gilliam, and other men of a certain age Grace Lee’s talk: Diary, Discourse and Demonetisation Star Wars is Over (in a good way) Election notes How to read articles on websites that don’t want you to read their articles Uranium Club RIP Tom Spurgeon Introducing Notes, my new microblog Lindelof’s Watchmen, and the creative employment of corporate IP Eisenhower and the Hippies The Widelux, as used by Jeff Bridges Media with Edges Art-trip to Liverpool Sunday Reads - spinning tops, Nazis, Icelandic art, microbiomes, Greta and the social rituals of maintenance. Birthday Books