Scaling a hippocampus model with GPU parallelisation and test-driven refactoring

Stevenson, Jack and Fox, Charles (2022) Scaling a hippocampus model with GPU parallelisation and test-driven refactoring. In: 11th International Conference on Biomimetic and Biohybrid Systems (Living Machines), 20-22 July 2022.

Documents
Scaling a hippocampus model with GPU parallelisation and test-driven refactoring
Accepted manuscript
[img]
[Download]
[img]
Preview
PDF
StevensonHCGPU.pdf - Whole Document

437kB
Item Type:Conference or Workshop contribution (Paper)
Item Status:Live Archive

Abstract

The hippocampus is the brain area used for localisation, mapping and episodic memory. Humans and animals can outperform robotic systems in these tasks, so functional models of hippocampus may be useful to improve robotic navigation, such as for self-driving cars.
Previous work developed a biologically plausible model of hippocampus based on Unitary Coherent Particle Filter (UCPF) and Temporal Restricted Boltzmann Machine, which was able to learn to navigate around small test environments. However it was implemented in serial software, which becomes very slow as the environments and numbers of neurons scale up. Modern GPUs can parallelize execution of neural networks.
The present Neural Software Engineering study develops a GPU accelerated version of the UCPF hippocampus software, using the formal Software Engineering techniques of profiling, optimisation and test-driven refactoring. Results show that the model can greatly benefit from parallel execution, which may enable it to scale from toy environments and applications to real-world ones such as self-driving car navigation. The refactored parallel code is released to the community as open source software as part of this publication.

Keywords:hippocampus, computational neuroscience, robotics, navigation, memory
Subjects:B Subjects allied to Medicine > B140 Neuroscience
Divisions:College of Science > School of Computer Science
ID Code:49936
Deposited On:24 Jun 2022 14:51

Repository Staff Only: item control page