Neural Network Fault Detection via Selective Layer Duplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fault detection methods in neural network processing are computationally intensive and costly, particularly in safety-critical systems, where they require full hardware duplication, which limits computational capacity and increases costs.
Innovation Solution
A method that schedules computations for neural network layers as either non-duplicated or duplicated operations, allowing for efficient fault detection by comparing outputs from duplicated computations, thereby reducing the need for full hardware duplication and enhancing computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full hardware duplication is used for fault detection, then reliability is improved, but device complexity and computational burden increase
Solution Approach 1:
The patent segments the neural network processing into different layers and identifies specific layers where fault detection is most critical. Instead of duplicating hardware for all layers, the invention applies fault detection mechanisms selectively to specific layers based on their importance and fault susceptibility, thereby reducing overall hardware complexity while maintaining reliability where it matters most.
Solution Approach 2:
The patent implements local quality by applying different fault detection strategies to different layers of the neural network. High-criticality layers receive duplicated computation and comparison-based fault detection, while lower-criticality layers use simpler or no fault detection mechanisms. This localized approach optimizes the balance between reliability and device complexity.
2Reliability
If full hardware duplication is used for fault detection, then reliability is improved, but productivity decreases
Solution Approach 1:
The patent divides the neural network into multiple layers and applies fault detection selectively to specific segments (layers) rather than the entire network. This segmentation allows critical layers to have robust fault detection while non-critical layers maintain high computational efficiency, thus preserving overall productivity.
Solution Approach 2:
The patent applies partial duplication of computations only where necessary for fault detection, rather than duplicating all computations. By performing duplicated computations partially (only for selected layers and operations), the system achieves adequate fault detection capability without the excessive computational burden of full duplication, thereby maintaining productivity.
3Measurement precision
If duplicated computations are performed for fault detection, then measurement precision is improved, but use of energy increases
Solution Approach 1:
The patent applies duplicated computations and comparison-based verification locally to specific layers and operations where fault detection is most beneficial, rather than uniformly across the entire neural network. This localized verification improves measurement precision for critical outputs while minimizing the additional energy consumption associated with duplicated computations.
Data Source
AI summary
A method of performing fault detection during computations relating to a neural network comprising a first neural network layer and a second neural network layer in a data processing system, the method comprising: scheduling computations onto data processing resources for the execution of the first neural network layer and the second neural network layer, wherein the scheduling includes: for a given one of the first neural network layer and the second neural network layer, scheduling a respective given one of a first computation and a second computation as a non-duplicated computation, in which the given computation is at least initially scheduled to be performed only once during the execution of the given neural network layer; and for the other of the first and second neural network layers, scheduling the respective other of the first and second computations as a duplicated computation.