Implementation of "Recurrent Models of Visual Attention" V. Mnih et al.
Modified from https://github.com/zhongwen/RAM https://github.com/jlindsey15/RAM
NOTICE: tf.stop_gradient should be applied to location sampled from Gaussian distribution parametered by (mean, std), so actor-critic reinforcement learning back propagate mean with (reward-tf.stop_gradient(baselines)) * loglikelihood to update loc network. Futhermore, baseline network is updated by (R-baseline)^2.
Run by python ram.py and it can reproduce the result on Table 1 (a) 28x28 MNIST