“If it doesn’t have any samples, or training data, or raw data to learn from then it can’t answer anything,” Jai says. “But if you have plenty of samples to learn from then it’s able to do the right thing.”
“The amount of errors it will make is directly proportional to the number of accurate pieces of data it can draw upon when it was trained,” Aidan adds.
When it comes to vomiting rainbows or wearing a flower crown, Snapchat does things a little differently. It relies on the similarities of faces to identify key features and apply animations when they move.
“Human faces have very little variance between them,” Aidan explains. “Once a feature is known, due to the similarities between people, it doesn’t need to be relearnt by looking at a new person it’s never seen before. It’s a simpler, more general purpose solution that doesn’t have to be trained other than to know what these key features look like.”