4 votes

Neuroevolution of Self-Interpretable Agents

1 comment

  1. DataWraith
    Link
    This is an incredibly cool and creative project. They take a flaw in human mental processing (attentional blindness) and make a virtue out of it by forcing the learning agent to pay attention to...

    This is an incredibly cool and creative project.

    They take a flaw in human mental processing (attentional blindness) and make a virtue out of it by forcing the learning agent to pay attention to no more than 10 image patches at once ("attention bottleneck"). This results in a neural network that has very few parameters, making it possible to evolve its weights with an evolutionary algorithm. And since it only pays attention to 10 different things, its 'thought process' can even be visualized.

    NOTE: Page has flickering images

    3 votes