Evolution of active categorical image classification via saccadic eye movement

I put together a couple demo videos for our Active Categorical Classifier (ACC) project that we'll be presenting at the PPSN 2016 conference.

If you're interested in this project and can't wait for PPSN, we have:

• a preprint of the paper on arXiv,
• the C++ code on GitHub, and
• a demo of the new Python version on GitHub.

In the first video, I show the entire image as the ACC roams around it trying to classify the object. Even though we can see the entire image at once, the ACC only sees the 9 pixels in its immediate vicinity.

In the second video below, I only show the pixels that the ACC has seen during the simulation. Can you guess what digit it is before the ACC?

I think the second video does a great job of showing how little information the ACC is processing to classify each image. Talk about a light-weight classifier!