Dissertation
Point-Neuron Modeling with post-von Neumann Experimental Neuromorphic System
Washington State University
Doctor of Philosophy (PhD), Washington State University
2023
DOI:
https://doi.org/10.7273/000004998
Abstract
Neuromorphic hardware is based on emulating the natural biological structure of the brain. Since its computational model is similar to many standard neural models, it could serve as a computational accelerator for research projects in the field of neuroscience and artificial intelligence. However, in order to exploit this new generation of computer chips, we ought to perform rigorous simulation and consequent validation of neuromorphic models against their conventional implementations. Intel’s fifth generation neuromorphic chip - “Loihi”, which is based on the idea of Spiking Neural Networks (SNNs) emulating the activity of neurons in the brain, serves as our neuromorphic platform. The goal of this thesis is to lay out a methodological and numeric groundwork to enable a comparison between neuromorphic and conventional platforms. The thesis tackles two central questions:
1. How precise are the Loihi simulations as compared to the classical simulations, given its distinct hardware and programming architecture?
2. How do different model parameters contribute to the robustness of the validated model?
The premise for answering the questions above is set with a focus on one of the most effective neural models, the Leaky Integrate-and-Fire (LIF) model. The data used for the simulations is based on neurons in the mouse primary visual cortex and matched to a rich data set of anatomical and physiological constraints. Simulations on classical hardware serve as the validation platform for the neuromorphic implementation. Thus, the questions are answered in two broad parts - (1) We establish a mapping that implements the model in Loihi by addressing the architectural and programming differences present in this novel platform. We find that with this mapping Loihi is able to replicate the classical simulations very efficiently with a correlation of ∼0.9999 while being much faster than BMTK with a magnitude of order ∼10^2. (2) In the second part, we analyze our cost function the Root Mean Square Error (RMSE), to examine the effect of the different model parameters. By extending the core perturbation and sensitivity analysis concepts, we characterize the error variation contributed by the different model parameters in addition to determining the most sensitive parameter, thus investigating the robustness of the model. We verify the mathematical results by enabling the perturbation and sensitivity simulations on both platforms and infer how network stimulus along with network architecture affect the results.
Metrics
4 File views/ downloads
46 Record Views
Details
- Title
- Point-Neuron Modeling with post-von Neumann Experimental Neuromorphic System
- Creators
- Srijanie Dey
- Contributors
- Alexander Dimitrov (Advisor)Bala Krishnamoorthy (Committee Member)Kevin Cooper (Committee Member)Andreas Wild (Committee Member)
- Awarding Institution
- Washington State University
- Academic Unit
- Mathematics and Statistics, Department of
- Theses and Dissertations
- Doctor of Philosophy (PhD), Washington State University
- Publisher
- Washington State University
- Number of pages
- 209
- Identifiers
- 99901019233301842
- Language
- English
- Resource Type
- Dissertation