Hybrid computing using a neural network with dynamic external memory

ABSTRACT

Artificial neural networks excel at sensory processing and pattern recognition but struggle with the systematic and reliable execution of algorithmic tasks. We introduce the Differentiable Neural Computer (DNC), a neural network architecture coupled with a read-write memory matrix that overcomes these limitations. The DNC can learn to use its memory to answer questions about complex, structured data, including navigating the London Underground and solving block puzzle tasks. By dynamically allocating memory and maintaining temporal links between stored vectors, the DNC achieves a level of systematic generalization previously unattainable by standard neural networks.