First Advisor
Christof Teuscher
Term of Graduation
January 2026
Date of Publication
9-10-2026
Document Type
Thesis
Language
English
Subjects
fpga, fractional-order, neuromorphic, reinforcement learning, spiking neural networks
Physical Description
1 online resource ( pages)
Abstract
Researchers working with Spiking Neural Networks (SNNs) are faced with the challenge of identifying useful biologically plausible features for inclusion in neuron models. Prior research indicates that dynamics of fractional-order calculus, which impart a form of intrinsic memory, are a key feature of biological neurons. In this work, we present the development and usage of neural networks with intrinsic memory in application to the cart-pole task, a common benchmark for neural network performance. We found that usage of fractional-order neurons may allow for up to a 37.5% reduction in neural network size as compared to networks using a standard Leaky Integrate-and-Fire (LIF) neuron. This reduction in network size can be leveraged to achieve efficiencies such as reduced latency or energy usage on inference tasks. In addition, fractional-order dynamics appear to be a particularly effective way to incorporate intrinsic memory, as we demonstrate that a method leveraging bitshifts instead of multiplication for inclusion of history required a higher number of neurons for learning the cart-pole task. Though we report challenges in obtaining consistent training results using a Deep Q-Networks (DQN) algorithm, our analysis of over 100 trials for each of three neural network types supports our conclusions regarding the trend in network sizes across models. The primary tradeoffs we identify for neurons with intrinsic memory are the additional resources to support calculation of a history term. We report these added requirements in the context of implementation and execution in a digital computing environment. Through our investigation, we created a software-to-hardware workflow for testing novel neuron models, beginning with training a network in software and resulting in execution on hardware. This work supports future research into custom neuron models, providing a pathway for evaluation of features, such as fractional-order dynamics, which may prove useful in improving accuracy and reducing network size. We have identified that fractional-order dynamics specifically show promise in enhancement of neural network performance, and this work may inspire future development of analog components with inherent fractional-order capabilities.
Rights
In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).
Recommended Citation
Anderson, Niklas Karl, "Intrinsic Memory in Digital SNNs: Enhancing Performance on the Cart-Pole Task with Fractional-Order Dynamics" (2026). Dissertations and Theses. Paper 7212.