Advanced Quantum Materials Lab · April – August 2025
Memristor Spiking Neural Networks
Research Overview
As an undergraduate researcher in the Advanced Quantum Materials Lab at Stevens, I worked on spiking neural networks (SNNs) built on memristor hardware. Conventional neural networks run as software on digital processors, and energy use is a limiting factor for AI training. In a memristor crossbar, computation happens in analog through the device physics, at lower energy cost.
Approach
I built and simulated memristor-based SNN architectures in PyTorch and Brian2, with device-level modeling so the simulation tracked how real memristors behave electrically. The main task was implementing and validating analog crossbar inference for spike-encoded digit classification.
- Frameworks: PyTorch for network architecture, Brian2 for spiking-neuron simulation.
- Device modeling: Precise device-level models of memristive behavior in the simulation loop.
- Task: Spike-encoded digit classification through analog crossbar inference.
Results
- Classification Accuracy: 92.63% on spike-encoded digit classification through the analog crossbar.
- Validation: Performance analysis and system validation using heatmaps and confusion matrices.
- Presentation: Findings presented at the Stevens Symposium.
Poster presented at the Stevens Symposium.