Selective N Modular Redundancy for Neural Network Fault Tolerance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional N Modular Redundancy (N-MR) techniques for neural networks are costly due to the need for replicating computations by a factor of N, either temporally or spatially, which is prohibitively expensive for generic systems and neural networks.
Innovation Solution
Implementing selective N modular redundancy (N-MR) by using small neural network checkers and applying it only to critical computations, reducing the cost by selectively replicating the neural network and applying N-MR at the neuron level based on criticality analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If N Modular Redundancy is applied to neural networks to improve fault tolerance, then reliability is improved, but computational cost increases by a factor of N
Solution Approach 1:
The patent segments the neural network into critical and non-critical computations, applying N-MR only to critical segments. This selective application reduces the overall computational cost while maintaining fault tolerance for the most important operations.
Solution Approach 2:
The patent implements local quality by differentiating between critical and non-critical computations, applying full redundancy only where necessary. This creates varying levels of fault tolerance across different parts of the network based on their importance.
2Reliability
If N Modular Redundancy replicates the neural network spatially through N redundant hardware units, then fault detection capability is improved, but device complexity increases
Solution Approach 1:
The patent segments redundancy application to only critical computations rather than replicating the entire network. This reduces the number of redundant hardware units needed while maintaining fault detection for important operations.
Solution Approach 2:
The patent applies partial redundancy by using N-MR selectively on critical computations rather than universally. This partial action approach reduces hardware complexity while providing sufficient fault detection where it matters most.
3Reliability
If N Modular Redundancy replicates computation temporally by repeating it N times, then fault tolerance is improved, but execution time increases
Solution Approach 1:
The patent segments the computational workflow to apply temporal repetition only to critical computations. This reduces overall execution time while maintaining fault tolerance for the most important operations.
Solution Approach 2:
The patent applies temporal redundancy partially, repeating computations only when necessary for critical operations. This partial application reduces the total time penalty while maintaining adequate fault tolerance.
Data Source
AI summary
An N modular redundancy method, system, and computer program product include a computer-implemented N modular redundancy method for neural networks, the method including selectively replicating the neural network by employing one of checker neural networks and selective N modular redundancy (N-MR) applied only to critical computations.