Related Experiment Video
Updated: Dec 14, 2025
07:34
A Simple Stimulatory Device for Evoking Point-like Tactile Stimuli: A Searchlight for LFP to Spike Transitions
Published on: March 25, 2014
10.2K
Supervised Learning in All FeFET-Based Spiking Neural Network: Opportunities and Challenges
Sourav Dutta1, Clemens Schafer2, Jorge Gomez1
1Department of Electrical Engineering, College of Engineering, University of Notre Dame, Notre Dame, IN, United States.
Frontiers in Neuroscience
|July 17, 2020
Summary
This study introduces an all-ferroelectric field-effect transistor (FeFET) based spiking neural network (SNN) hardware for energy-efficient neuromorphic computing. The FeFET SNN platform enables in-memory computing and supports surrogate gradient learning for machine learning tasks.
Area of Science:
- Neuromorphic Engineering
- Artificial Intelligence
- Materials Science
Background:
- Spiking Neural Networks (SNNs) offer potential for energy-efficient AI by mimicking brain computation with spikes, deviating from traditional deep learning.
- Current SNN hardware often relies on CMOS technology and segregated computation/memory, limiting energy and areal efficiency.
- Ferroelectric Field-Effect Transistors (FeFETs) show promise for neuromorphic hardware due to their non-volatile nature and polarization dynamics.
Purpose of the Study:
- To propose and demonstrate an all-FeFET-based Spiking Neural Network (SNN) hardware.
- To achieve low-power, spike-based information processing with co-localized memory and computing (in-memory computing).
- To implement a surrogate gradient (SG) learning algorithm for supervised learning on the FeFET SNN platform.
Main Methods:
- Experimental demonstration of neuronal and synaptic dynamics using 28 nm high-K metal gate FeFET technology.
- Implementation of a surrogate gradient (SG) learning algorithm for supervised learning on the developed SNN hardware.
- Device-algorithm co-design considering device-level variations and limited bit precision for synaptic weights.
Main Results:
- Successful experimental validation of essential neuronal and synaptic functionalities in FeFETs for SNNs.
- Demonstration of supervised learning on the MNIST dataset using the SG algorithm on the FeFET SNN platform.
- Analysis of the impact of device stochasticity and limited analog states on classification accuracy.
Conclusions:
- An all-FeFET-based SNN hardware platform is presented, enabling energy-efficient neuromorphic computing and in-memory processing.
- The platform successfully supports traditional machine learning algorithms through surrogate gradient learning.
- Synergistic device-algorithm co-design is crucial for optimizing the performance of FeFET-based neuromorphic hardware.
