SPACE10 + CIRG: AI Aerobics

The experience

AI Aerobics is a digital experience where you follow a machine’s instructions on how to move. Sounds simple enough, right? In fact, the poses the computer demands of you reflect what it’s learned about unique movement through reinforcement learning techniques. And not only that: they’ve also been tweaked in response to what CIRG found out about movement through talking to choreographers, yoga teachers and more. In the experiment itself, however, you don’t get that whole backstory. Instead, you get an entertaining experience where you work out like a machine says you should—and then export your routine as a GIF if you wish to.

How we got there

The process of teaching machines to generate novel movement and having them teach us in return was, perhaps unsurprisingly, multi-layered.

Step 1: Moving like a machine

First, CIRG used reinforcement learning to enable the machine — depicted as a mannequin — to generate completely randomised movements. The point here was to avoid the computer mirroring our own human biases or cultural knowledge in terms of how we move; but just as importantly, we wanted our AI to move like a computer, not like a person. ‘The only data the AI had was about its own joints and the environment—how gravity works, for example,’ says Jochen Weber, co-founder of CIRG. ‘It didn’t necessarily know that an arm is an arm. It didn’t know that you have to move your hips when you move. Reducing the information provided to the bare minimum enabled the AI to come up with its own idea of devising a body, keeping it away from any data set around how we move.

Step 2: Searching for the ideal movement

Once our AI had learned how to move, CIRG began the second part of the experiment—interviewing people who specialise in movement, like choreographers and physiotherapists and integrating that knowledge into our machine’s movements.

In a way, the purpose was to end up with an AI that reflects a whole spectrum of democratically ‘ideal’ ways of moving. Indeed, CIRG wanted to extract a wide body of knowledge about movement—from the practical to the artistic, from the cultural to the automatic—and embed those values in our AI. ‘Within that spectrum, the physiotherapist is about optimising movement, for example,’ says Weber. ‘The physiotherapist wants you to unlearn movements and replace them with the right movements as precisely as possible. On the other side, choreographers are about interpreting movement for your own body and purpose.’

So, how did we teach this knowledge to our AI mannequins? What did CIRG choose to incorporate, and what information did they choose to leave out? The starting point was showing the interview subjects how our mannequins moved—and then paying attention to what they themselves found particularly curious. ‘The physiotherapist saw that the mannequins were using their upper bodies a lot. That’s something we as people don’t do a lot because of our cultural habits,’ explains Weber. ‘We usually use our hips a lot, but our body is actually not really made for that.’

Once CIRG had enough data of this nature from the interviews—learning about interpretations of ideal movement as well as what our mannequins are potentially already doing better or differently than humans—they began to edit our AI’s movements to reflect these learnings. And voila: an AI Aerobics instructor, imbued with its own unbiased learnings but also reflections from movement professionals, was born. And now that it’s learned from itself and from us, the tables turn: it’s time for us humans to learn from our AI Aerobics instructor through a digital experience.

Why is this important?

Today, biased AI is quietly becoming the norm. From recommender algorithms which show us goods or services which reinforce our tastes and our viewpoints, to corporate recruitment tools which won’t select women and minorities for interviews—the machines we create often reflect human prejudices. And, if we consider how wide-reaching and ever-present AI is becoming, those subtly communicated biases will come with sinister consequences.

Unless we make active efforts to enable these machines to remain blind to sociocultural biases, that is. For us, this mission is the core of our experiment. While the end result is a lighthearted experience about physical movement, the process we took still proves a promising thing is possible: we can create an AI that’s protected from much of our human knowledge—a machine that comes to its own, unique conclusions about the world.

Once that happens, a different vision for the future of machine learning opens up. Could AI be beneficial to us without being intrusive? Could we learn new skills from AI while maintaining our autonomy? Because in this experiment, we consciously choose to engage with our AI Aerobics instructor. We can mimic the movements for however long we want, and exit the experiment when we’ve had enough.

In short: instead of remaining ignorant to just how prevalent AI is in our lives, we engage in a transparent relationship and actually set the rules of engagement. We maintain our human autonomy while also learning from how a computer understands the world. And isn’t that the future of AI and machine learning we’d want to believe in?