google-deepmind/dnc
A TensorFlow implementation of the Differentiable Neural Computer, a neural network architecture augmented with differentiable external memory.

This repository provides an implementation of the Differentiable Neural Computer published in Nature by DeepMind. It implements the DNC as a collection of composable RNNCore modules, including an access module that handles memory read/write operations using temporal linkage and freeness tracking, and a controller module that determines memory access patterns. The architecture augments a neural network with an external memory matrix that can be read from and written to in a differentiable manner, enabling the model to solve complex relational reasoning tasks.