Brain-Inspired Physical Reservoir Computing Architectures Using Biomolecular Memristors
Open Access
- Author:
- Armendarez, Nicholas
- Graduate Program:
- Mechanical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- February 20, 2025
- Committee Members:
- Robert Kunz, Professor in Charge/Director of Graduate Studies
Herschel Pangborn, Major Field Member
Joseph Najem, Chair & Dissertation Advisor
Jean-Michel Mongeau, Major Field Member
Abhronil Sengupta, Outside Unit & Field Member - Keywords:
- Nonlinear Dynamical Systems
Reservoir Computing
Memristors - Abstract:
- Brain-inspired computing has primarily driven advancements in machine learning algorithms and hardware. The low energy consumption and exceptional classification and prediction capabilities of the human brain make it an intriguing model for computation and information processing. However, most research has focused on software implementations of neural networks, which demand immense computational power for training and operation. At its core, neural network research hinges on connections between nonlinear nodes that map input data into a high-dimensional space for easier categorization. Traditional computing executes these nonlinear functions digitally, storing connections as weights in memory that the processor must retrieve to perform calculations. In contrast, biological brains integrate nonlinear activation, evident in the analog dynamics of neurons and synapses, along with massively parallel connections between these computational units where memory and processing are co-located. Thus, the materials and methods of contemporary computing are incompatible with achieving brain-like computing. To address this challenge, this dissertation explores biological materials and assemblies for computation and information processing, specifically ion channels embedded in phospholipid membranes, and their application in neural network architectures. The lipid bilayer and its voltage-gated ion channels form the fundamental substrate of biological computation, enabling efficient information processing. The work first focused on investigating the properties and nonlinear dynamics of two-terminal biomolecular memristors made from droplet interface bilayers doped with ion channels, modeling their behavior using both linear and nonlinear first-order models and identifying the conditions under which each model applies. Next, these biomolecular memristors’ inherent nonlinearities and dynamic memory were leveraged for reservoir computing, a recurrent neural network paradigm. Reservoir computing relies on a sparsely connected, randomly generated reservoir layer with nonlinear nodes, and recent research has demonstrated that various physical systems can serve as such reservoirs in analog form. In this work, ion channel devices were demonstrated to function as parallel, low-dimensional reservoirs for diverse classification tasks. Additionally, by employing a nonlinear memristor model, physical reservoir layers were trained ex-situ and successfully performed in-situ inference without significant accuracy loss. State-of-the-art parallel memristor reservoir computing architectures lack recurrence among nodes—a crucial property for effective computation. To overcome this, the unique attributes of biomolecular memristor assemblies were leveraged to introduce recurrence for the first time. Networks of these ion channel devices were constructed to generate high-dimensional dynamics capable of solving complex problems requiring substantial memory and nonlinearity. Ultimately, nonlinear networks exhibiting fading memory dynamics were developed. Through accurate modeling and simulations, it is demonstrated that coupled networks of memristive devices can significantly outperform existing parallel implementations in the literature.