Now that’s just unfair to former Webster star Emmanuel Lewis.
2 Likes
{looks in mirror}
.
.
.
.
.
{pokes out Harrison Ford’s eyeballs}
.
{flees}
2 Likes
Wait, people with a coloboma are fake??
1 Like
Because the AI has no idea it’s generating a face.
Someone trained the network by feeding photos of faces; then said “your turn. Make an array of ones and zeros similar to the ones I just gave you.” The algorithm produced output, then the human said “yeah, good” or “no, not so good .” Repeat a thousand times and it gets better and better.
Whatever those thousand lessons contained was absorbed into its digital neurons as arrays of numbers, and not as distinct features of a face. It wasn’t told anything like “expect a face where a face has a nose in the middle with two eyes and a mouth”, or that “an eye has a white round orb with a colored ring around a black circle”, etc. It’s just a collection of numbers that we know to have patterns; the neural network eventually learns what patterns make for good output through training and feedback.
Here’s a slide deck with really cool images explaining what’s going on inside the process. Warning: many of the intermediate images it produces look like a cruel, disturbed Picasso assembled a dog. If you are triggered by photos of spider-like collies with three eyes, well, you were warned.
3 Likes
This is also why simple “adversarial perturbations” cause AI image recognition systems to misidentify images in ways that have nothing to do with the shape of the object.
3 Likes
Perhaps it was trained on photos of people with syphilis.
This topic was automatically closed after 5 days. New replies are no longer allowed.
