Enter the maze

Breeding salience

Some of Milan's test patterns

Milan Verma from Queen Mary has created a computer model of the way humam's see. It particularly concerns whether we notice things in a scene or not. How to test such a model though? He decided to generate test patterns for both humans and his program to look at that he creates using an Artificial Intelligence program.

Milan's work is the first time that artificial intelligence has been used to create patterns with such precise differences in salience to test human perception. His program uses a form of artificial intelligence called a genetic algorithm. It is a process based on Darwin's ideas of natural selection and the 'survival of the fittest' that drives evolution.

The artificial intelligence first generates a set of random patterns. In nature, 'survival' means having children before you die. For a test pattern the survival test is about how close the model rates it to the given target value of salience. Patterns survive if, from all those currently competing, they have values judged closest to the target. The others are killed off. New patterns (children) are created by making random changes (mutations) to the survivor patterns. The children are then tested for fitness and the process continues.

Keep doing this over and over and all the small but successful mutations build up, all the while getting closer to the desired level of salience. The result is a box that pops out exactly the right amount. In the patterns above the model predicts that a should pop out more than b, which should be more obvious than c and that more than d.

The artificial intelligence breeds never-before-seen patterns to a custom-set level of difficulty. For some images it's easy to spot the difference. In other images it is much harder.