Last week, I ran five selected snippets through twenty different image recognition systems, completing one hundred iterations in total. What came back were sixty-six distinct types of output — sixty-six different ways of defining the same five images.
And then I noticed something really interesting. When I zoomed into these expressions, YOLO started identifying them as some pretty unexpected things. So I asked AI about it, and it said it was probably because of the eye shape — a round form with a black ring and a white centre looks, to the model, like a hollow structure. So it classified it as a cup. And I found that really fascinating. Because in the world of these models, there’s no such thing as understanding an image. To them, an image is just a formula — a combination of features added together.
So I started wondering — what if I deliberately disrupted those features? What would the model do? I started by manually extracting the individual features of a panda image, and then I used a difference function to invert and reassemble them. I kept the colour information because I felt like colour is still a pretty important factor in how machines read images.
But when I was doing it manually, I kept catching myself making unconscious decisions. Like I would instinctively avoid breaking up the crescent shape because it looked like a moon and I didn’t want to lose that. So I realised that as long as I was the one making the decisions, the process wasn’t really objective. So I decided to hand it over to a program — randomise the order, let the machine do the disrupting. And I built a website around that.
And this is the process of building the website.
This is what the website looks like. The middle column shows the randomly generated result each time. When you click the recombine button, the SVG paths on the left update in real time, the right side accumulates every version that’s been generated, and the tags at the top build up as a record of every label the image has received.
Website Video:https://youtu.be/eIVKxvRTRus
And this is the poster I made by pulling all the results together.
Leave a Reply