Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Testing Spintronics Implemented Monte Carlo Dropout-Based Bayesian Neural Networks

Soyed Tuhin Ahmed, Michael Hefenbrock, Guillaume Prenat, Lorena Anghel, Mehdi B. Tahoori

TL;DR

This work tackles the reliability and testing of dropout-based Bayesian neural networks deployed on Spintronics-CIM, where stochastic dropout and hardware non-idealities challenge deterministic functional testing. It models non-idealities, introduces a repeatability ranking-based automatic test pattern generation framework, and develops a lightweight online fault-detection method that leverages a Gaussian uncertainty distribution with bounds $\mu \pm 3\sigma$. The authors demonstrate near-complete fault coverage for critical faults and conductance variations across SpinDrop, SpatialSpinDrop, and ScaleDrop on CIFAR-10 with ResNet-18, while requiring only $0.2\%$ of training data as test vectors. The proposed approach achieves high fault-detection efficiency and low false alarm rates, with detailed overhead and scalability analyses, offering a practical pathway for safe, in-field testing of BayNNs in spintronic hardware.

Abstract

Bayesian Neural Networks (BayNNs) can inherently estimate predictive uncertainty, facilitating informed decision-making. Dropout-based BayNNs are increasingly implemented in spintronics-based computation-in-memory architectures for resource-constrained yet high-performance safety-critical applications. Although uncertainty estimation is important, the reliability of Dropout generation and BayNN computation is equally important for target applications but is overlooked in existing works. However, testing BayNNs is significantly more challenging compared to conventional NNs, due to their stochastic nature. In this paper, we present for the first time the model of the non-idealities of the spintronics-based Dropout module and analyze their impact on uncertainty estimates and accuracy. Furthermore, we propose a testing framework based on repeatability ranking for Dropout-based BayNN with up to $100\%$ fault coverage while using only $0.2\%$ of training data as test vectors.

Testing Spintronics Implemented Monte Carlo Dropout-Based Bayesian Neural Networks

TL;DR

This work tackles the reliability and testing of dropout-based Bayesian neural networks deployed on Spintronics-CIM, where stochastic dropout and hardware non-idealities challenge deterministic functional testing. It models non-idealities, introduces a repeatability ranking-based automatic test pattern generation framework, and develops a lightweight online fault-detection method that leverages a Gaussian uncertainty distribution with bounds . The authors demonstrate near-complete fault coverage for critical faults and conductance variations across SpinDrop, SpatialSpinDrop, and ScaleDrop on CIFAR-10 with ResNet-18, while requiring only of training data as test vectors. The proposed approach achieves high fault-detection efficiency and low false alarm rates, with detailed overhead and scalability analyses, offering a practical pathway for safe, in-field testing of BayNNs in spintronic hardware.

Abstract

Bayesian Neural Networks (BayNNs) can inherently estimate predictive uncertainty, facilitating informed decision-making. Dropout-based BayNNs are increasingly implemented in spintronics-based computation-in-memory architectures for resource-constrained yet high-performance safety-critical applications. Although uncertainty estimation is important, the reliability of Dropout generation and BayNN computation is equally important for target applications but is overlooked in existing works. However, testing BayNNs is significantly more challenging compared to conventional NNs, due to their stochastic nature. In this paper, we present for the first time the model of the non-idealities of the spintronics-based Dropout module and analyze their impact on uncertainty estimates and accuracy. Furthermore, we propose a testing framework based on repeatability ranking for Dropout-based BayNN with up to fault coverage while using only of training data as test vectors.
Paper Structure (24 sections, 7 figures)

This paper contains 24 sections, 7 figures.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Dropout-Based Bayesian Neural Networks
  4. Spintronics Device and Bayesian Inference in Spintronics-CIM
  5. Related Works
  6. Failure Mechanisms of BayNN in Spintronics-CIM
  7. Failure Mechanisms of Spintronics Device
  8. Failure Mechanisms of Buffer Memories
  9. Failure Mechanisms of Spintronic-Based Dropout Module
  10. Proposed Approach
  11. Problem Statement
  12. Automatic Test Generation Framework
  13. Proposed Fault Detection Approach
  14. Reduction of False Positives Rate
  15. Evaluation
  16. ...and 9 more sections

Figures (7)

  • Figure 1: Conventional and Monte Carlo Dropout-based Bayesian inference, where the same input is applied to the network with T times with T different Dropout configurations to get the approximate posterior distribution.
  • Figure 2: Stochasticity of a) Spintronics-CIM output (logits values) and b) uncertainty estimates for the same input for $200$ different predictions and inference runs, respectively.
  • Figure 3: Impact of Inference accuracy of Spintronics BayNNs with different bit-flip and stuck-at-fault rates, compared to a baseline without faults.
  • Figure 4: Comparison of the impact of inference accuracy of Spintronics BayNNs with conductance variations relative to a fault-free baseline.
  • Figure 5: Fault coverage of proposed approach on various Spintronics implemented BayNN methods under varying bit-flip and stuck-at faults rate a) affecting Spintronics cells that store weights, b) buffer memories that store intermediate activation, and c) different conductance variations in Spintronics.
  • ...and 2 more figures