Neural modeling of multiple memory systems and learning
Date of Issue2015
School of Computer Engineering
Centre for Computational Intelligence
This thesis presents a biologically inspired multi-memory system for modeling the structures and connections between the procedural and declarative memories. Using multi-channel self-organizing neural networks as building blocks, the proposed multi-memory system includes a procedural memory model that learns decision through reinforcement learning, an episodic memory model that encodes an individual's experience in the form of events and their spatio-temporal relations, and a semantic memory that captures factual knowledge. We have further proposed two major interaction process between the three memories. We further investigated the overall performance of the memory system on a first person shooting game and a Starcraft Broodwar strategic game. Our experimental results show that the system model is able to learn various forms of knowledge for the different domain tasks. The results also confirms that the memory interaction can lead to a significant improvement in both learning efficiency and performance.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence