Neuromorphic Fault Screening Using Misclassification-Driven Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning models mapped to neuromorphic hardware like memristor-based crossbars face reliability issues due to manufacturing defects and process variations, with critical faults causing misclassification, while benign faults are prevalent, leading to inefficiencies in fault recovery and hardware robustness.
Innovation Solution
A misclassification-driven training (MDT) method identifies critical faults by perturbing neural network parameters, using a gradient-based algorithm to pinpoint significant deviations, and trains a machine learning model to classify fault criticality, enabling targeted fault recovery and reducing overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fault recovery techniques are applied to all faults in neuromorphic hardware, then reliability is improved, but processing overhead and hardware wear-out increase due to the prevalence of benign faults
Solution Approach 1:
The patent applies different quality levels of fault recovery to different locations in the hardware. Critical faults (affecting less than 1% of operations) receive full fault recovery treatment, while benign faults (affecting more than 1% of operations) are allowed to propagate. This local differentiation resolves the contradiction by concentrating recovery resources only where they provide meaningful reliability improvement, eliminating wasted overhead on benign faults.
Solution Approach 2:
The patent segments the fault space into critical faults and benign faults based on their impact on inferencing accuracy. By segmenting rather than treating all faults uniformly, the system achieves reliability improvement for critical faults while avoiding the processing overhead associated with treating benign faults, thus resolving the contradiction between reliability and time loss.
2Reliability
If fault recovery is performed for all detected faults, then reliability is improved, but hardware wear-out increases due to repeated recovery operations
Solution Approach 1:
The patent applies fault recovery selectively based on fault criticality rather than universally. By identifying that less than 1% of faults are critical and applying recovery only to those locations, the hardware lifetime is preserved by avoiding unnecessary recovery operations on benign faults, while reliability is maintained through targeted recovery of critical faults.
Solution Approach 2:
Instead of applying full fault recovery to all faults (excessive action), the patent applies partial action only to critical faults. This partial approach is sufficient to maintain reliability while avoiding the wear-out caused by excessive recovery operations on benign faults, thus extending hardware lifetime.
3Reliability
If comprehensive fault detection is performed, then reliability is improved, but processing overhead increases due to the need to analyze all faults
Solution Approach 1:
The patent extracts and identifies only the critical subset of faults (less than 1% of total faults) that actually impact reliability. By taking out the essential critical faults from the complete fault set and focusing detection efforts only on those, the system achieves high detection accuracy for reliability-critical issues while maintaining processing efficiency by ignoring benign faults.
Solution Approach 2:
The patent segments the fault detection task into identifying critical faults versus benign faults. By segmenting the problem space and focusing computational resources only on identifying the critical 1% of faults that matter for reliability, the system achieves effective fault detection without the processing overhead of comprehensive analysis of all faults.
Data Source
AI summary
The disclosure provides misclassification-driven training (MDT) that efficiently identifies critical faults in neuromorphic hardware, such as a memristor crossbar. MDT advantageously identifies whether a hardware fault is a critical fault and can be used to limit fault recovery when a hardware fault is not a critical fault. By applying fault-tolerant techniques directed to critical faults, such as only for critical faults, processing overhead of a neural network can be reduced. In one aspect, the disclosure provides a method of identifying critical faults in neuromorphic hardware of a neural network. In one example the method of identifying includes: (1) determining a significant parameter of a trained neural network that impacts classification of a sample of a dataset, (2) obtaining a location of the significant parameter in the neuromorphic hardware, and (3) identifying the location as a critical fault of the neuromorphic hardware.