Topic: PredRNN

Visually understanding the world requires us to interpret surface properties like shape, depth, and reflectance from retinal images—with little or no access to the ground truth about these properties from which to learn. Previous work showed that perception and misperception of surface reflectance in static images is well predicted by unsupervised learning in a feedforward PixelVAE network. Here we extend into the temporal domain, testing whether learning-by-prediction in a recurrent network discovers the properties of moving surfaces to which humans are sensitive. We rendered 10,000 close-up