Linear map: sum_i(args[i] * W[i]), where W[i] is a variable. Args: args: a 2D Tensor or a list of 2D, batch x n, Tensors. output_size: int

Adds gradient noise as described in http://arxiv.org/abs/1511.06807 The input Tensor `t` should be a gradient. The output will be `t` + gaussi

We could have used RNN for parsing sentence but that tends to overfit. The simpler choice would be to take sum of embedding but we loose loose pos

Linear softmax answer module

Calculate loss

add data placeholder to graph

Calculate and apply gradients

Generate episode by applying attention to current fact vectors through a modified GRU

Use question vector and previous memory to create scalar attention for current fact

Get fact (sentence) vectors via embedding, positional encoding and bi-directional GRU

Get question vectors via embedding and GRU

Performs inference on the DMN model

Loads train/valid/test data and sentence encoding

Attention GRU with nunits cells.