Date: | 2020, November 3 |
Time: | 09:00 a. m. |
Place: | Online |
Author: | Al Amin, Raahim |
Title: | Memory Augmented Neural Network as an Accelerator for Graph Data Processing |
Artificial Neural Networks perform remarkably well to model complex patterns and prediction
problems. However, these models fall short in performance when it comes to store data over
a long period of time and have dependencies among them due to lack of memory.
Differentiable Neural Computer (DNC) was introduced by DeepMind researchers which is a
memory augmented neural network with an external memory. This network can perform read
or write operation in the memory matrix rather than random-access memory in traditional
computers. The authors showed that this network can be trained in graph data and can be
used in graph related queries for example, finding the shortest path between the two nodes of
a graph.
In spite of having several algorithms to find the shortest path between two nodes, we wanted
to see if DNC can be used as an accelerator for answering graph queries. In today’s
presentation, the implementation and results will be discussed to show whether DNC preforms
better than a traditional graph database management system called Neo4j. The performance
of the two systems is discussed based on the accuracy of the result of the shortest path query
and runtime.