DRAM

A fully differentiable recurrent visual attention model to classify MNIST digits. Some code / ideas taken from this implementation (https://github.com/ericjang/draw) of this paper (http://arxiv.org/pdf/1502.04623.pdf). Used in particular to examine the effect of pretraining on a pixel reconstruction task before training for classification.