“Learning to See” is a 2017 interactive installation by the Turkish-British artist Memo Akten in which neural networks analyze a live camera feed pointed at a table of everyday objects. As a visitor rearranges a cloth or a cable, the system reinterprets the scene in real time, redrawing it as oceans, clouds, fire, or flowers - whatever the network was trained to expect. The machine cannot see the objects as they are; it can only map them onto what it already knows.
Akten uses the piece to make a point about perception that applies to people as much as machines. In his own words, the work is built around the idea that “we see things not as they are, but as we are.” Technically he describes it as combining “Custom software, Artificial Intelligence, Machine Learning, Deep Learning, Generative Adversarial Networks.” Different versions of the network, trained on different image sets, are swapped in so the same physical scene produces wildly different hallucinated worlds.
The result is both beautiful and unsettling, an early artwork that treats a generative model’s limitations and biases as the subject rather than a flaw to hide.
Why business readers should care: Learning to See made the abstract idea of model bias tangible - a system projects its training onto whatever it is shown - years before that same dynamic became a practical concern in deployed AI products.