[Paper Review] Alleviation of Temperature Variation Induced Accuracy Degradation in Ferroelectric FinFET Based Neural Network

Sourav De, Yao‐Jen Lee+4 more|arXiv (Cornell University)|Mar 3, 2021

Ferroelectric and Negative Capacitance Devices10 citations

TL;DR

This paper investigates temperature-induced accuracy degradation in all-ferroelectric FinFET-based neural networks and proposes read voltage optimization in binary neural networks to maintain inference accuracy across temperature variations. By modeling conductance drift and optimizing read voltages, the approach preserves 96% accuracy at 233 K, mitigating degradation observed in baseline models trained at 300 K.

ABSTRACT

This paper reports the impacts of temperature variation on the inference accuracy of pre-trained all-ferroelectric FinFET deep neural networks, along with plausible design techniques to abate these impacts. We adopted a pre-trained artificial neural network (N.N.) with 96.4% inference accuracy on the MNIST dataset as the baseline. As an aftermath of temperature change, a compact model captured the conductance drift of a programmed cell over a wide range of gate biases. We observed a significant inference accuracy degradation in the analog neural network at 233 K for an N.N. trained at 300 K. Finally, we deployed binary neural networks with "read voltage" optimization to ensure immunity of N.N. to accuracy degradation under temperature variation, maintaining an inference accuracy of 96%. Keywords: Ferroelectric memories

Motivation & Objective

• Address the challenge of temperature variation-induced accuracy degradation in ferroelectric FinFET-based neural networks.
• Investigate the impact of temperature changes on conductance drift in ferroelectric memory cells used in neural network inference.
• Develop design techniques to ensure robustness of neural network inference under varying thermal conditions.
• Maintain high inference accuracy in binary neural networks despite temperature fluctuations, particularly at low temperatures.
• Demonstrate the effectiveness of read voltage optimization in achieving temperature-immune neural network performance.

Proposed method

• Utilized a pre-trained artificial neural network with 96.4% accuracy on MNIST as the baseline model.
• Developed a compact model to capture conductance drift of ferroelectric FinFET cells across a wide range of gate biases and temperatures.
• Simulated inference accuracy degradation at 233 K for a network trained at 300 K, identifying temperature sensitivity as a key issue.
• Implemented binary neural networks to reduce complexity and enhance robustness to device variations.
• Optimized the read voltage in binary neural networks to counteract temperature-induced conductance drift and maintain accuracy.
• Validated the proposed method by measuring inference accuracy across temperature ranges, focusing on stability at low temperatures.

Experimental results

Research questions

• RQ1 How does temperature variation affect the inference accuracy of all-ferroelectric FinFET-based neural networks?
• RQ2 What is the extent of conductance drift in ferroelectric FinFET cells across different gate biases and temperatures?
• RQ3 Can read voltage optimization in binary neural networks mitigate temperature-induced accuracy degradation?
• RQ4 To what extent can inference accuracy be preserved across temperature variations using the proposed method?
• RQ5 How does the performance of a neural network trained at 300 K degrade at 233 K without compensation techniques?

Key findings

• A significant inference accuracy degradation was observed in the baseline neural network at 233 K, despite being trained at 300 K.
• The compact model accurately captured conductance drift of ferroelectric FinFET cells over a wide range of gate biases and temperatures.
• Read voltage optimization in binary neural networks successfully mitigated temperature-induced accuracy degradation.
• The optimized binary neural network maintained an inference accuracy of 96% at 233 K, demonstrating robustness to temperature variation.
• The proposed method achieved near-baseline performance (96.4%) at low temperature, indicating strong immunity to thermal fluctuations.
• Conductance drift across temperature variations was identified as a primary cause of accuracy degradation in analog neural networks using ferroelectric devices.