In-Situ ANN Retraining for Memory Error Compensation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Memory subsystems deployed in harsh environments, such as remotely deployed sensors, face challenges with reliability and mean time to failure due to aging and environmental changes, leading to potential errors in artificial neural networks (ANNs) that can bias critical computations.
Innovation Solution
In-situ retraining of ANNs using a representative dataset and known outputs to adapt to minor errors in memory, allowing the system to extend its lifetime and reliability without relying on redundant memory solutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If redundant memory solutions are used to improve reliability, then system reliability improves, but device complexity and cost increase
Solution Approach 1:
The ANN performs self-retraining using its own stored weights and a representative dataset, adapting to memory errors autonomously without external intervention or redundant components. The system serves its own repair needs by evaluating its performance and retraining when errors are detected
Solution Approach 2:
The system changes the parameters of the ANN weights through retraining to adapt to memory errors. By periodically retraining the ANN with the representative dataset, the weights are adjusted to compensate for bit errors in the memory cells, maintaining accuracy without adding redundancy
2Reliability
If redundant memory solutions are used to improve reliability, then system reliability improves, but manufacturing cost increases
Solution Approach 1:
The existing memory and ANN infrastructure is used to perform self-diagnosis and self-repair through autonomous retraining, eliminating the need for expensive redundant memory components or external repair mechanisms
Solution Approach 2:
Instead of investing in expensive redundant memory that may never be used, the system uses inexpensive periodic retraining operations to compensate for memory degradation, treating the retraining as a low-cost maintenance operation
3Measurement precision
If ANN is retrained periodically to maintain accuracy, then ANN accuracy is maintained, but loss of time occurs due to retraining operations
Solution Approach 1:
The ANN is retrained periodically based on performance evaluation results rather than continuously. The controller evaluates the ANN at defined intervals and only triggers retraining when performance degradation is detected, optimizing the balance between accuracy maintenance and time loss
Solution Approach 2:
The system uses performance evaluation feedback to determine when retraining is necessary. By monitoring ANN performance metrics and comparing them against thresholds, the system triggers retraining only when needed, avoiding unnecessary time loss from routine retraining operations
Data Source
AI summary
An artificial neural network can be allocated to memory and operated. Performance of the artificial neural network can be periodically evaluated. The evaluation can include inputting a representative dataset to the artificial neural network and comparing an output of the artificial neural network to a known output for the representative dataset. The artificial neural network can be retrained at least partially in response to the evaluation yielding a sub-threshold result.