You might have seen a recent publication about how seemingly nonsense images fooled a neural network. Heres the comment that made the most sense to me to explain this. Few months ago, Google released their research showing the idea that they could change the school bus to not a school bus, by changing imperceptible things. They then used their perturbation algorithm to automatically train the network. I wondered if Jeff Hawkins idea is right. I wonder if you need the network to be able to rotate and shift the image slightly, in order to create a time varying representation. When humans look at a picture, the picture does not remain still. Our eyes move over the picture. Our head bobbles, changing the orientation. I imagine that if an image were directly placed on our rods and cones that we might have the same problems. Area for future research.
Posted on: Tue, 23 Dec 2014 09:09:22 +0000