Artificial neural networks are remarkably adept at sensory processing, sequence learning and reinforcement learning, but are limited in their ability to represent variables and data structures and to store data over long timescales, owing to the lack of an external memory.
Here we introduce a machine learning model called a differentiable neural computer (DNC),
We aim to combine the advantages of neural and computational processing by providing a neural network with read–write access to external memory.
A DNC is a neural network coupled to an external memory matrix.
The read vector r returned by a read weighting $w^r$ over memory M is a weighted sum over the memory locations:
Similarly, the write operation uses a write weighting $w^w$ to first erase with an erase vector $e$, then add a write vector $v$:
$M[i,j]$ ← $M[i,j](1 − w^w[i]e[j]) + w^w[i]v[j]$.
The functional units that determine and apply the weightings are called read and write heads.
The heads use three distinct forms of differentiable attention.
Our first experiments investigated the capacity of the DNC to perform question answering.
We therefore turn next to a set of synthetic reasoning experiments on randomly generated graphs.
After training with curriculum learning using graphs and queries with gradually increasing complexity, the networks were tested (with no retraining) on two specific graphs as a test of generalization to realistic data:
For the traversal task (Fig. 2b), the network was instructed to report the node arrived at after leaving a start node and following a path of edges generated by a random walk.