Neural Network Self-Correction via Boundary Condition Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks lack the ability to effectively recognize and correct incorrect output signals and are unable to adaptively reduce mistakes with new information, leading to ineffective self-correction mechanisms.
Innovation Solution
A controller system that includes a first neural network trained with initial data and a neural network manager to re-train the network incrementally using new data from sensors, with boundary conditions to filter and store input-output data for later training, and a security architecture to prevent system malfunctions and virus attacks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a conventional neural network is trained with initial data, then it can generate output signals, but it cannot recognize or correct incorrect output signals
Solution Approach 1:
The patent implements a feedback mechanism where the neural network's output is monitored by a boundary condition monitoring module that detects incorrect outputs and triggers re-training with new data, enabling the system to self-correct and improve reliability over time
Solution Approach 2:
The system performs preliminary actions by collecting and storing new data in boundary conditions before incorrect outputs occur, preparing training data in advance so that when errors are detected, the network can immediately re-train with relevant new information
2Productivity
If the neural network operates continuously without monitoring, then productivity is maintained, but incorrect outputs cannot be detected or corrected
Solution Approach 1:
A monitoring module continuously observes the neural network's output signals and provides feedback when boundary conditions indicate incorrect outputs, enabling error detection without interrupting overall system productivity
Solution Approach 2:
The boundary condition monitoring module acts as an intermediary between the neural network and the re-training process, detecting errors and triggering corrective actions while allowing the main system to continue operating
3Measurement precision
If re-training is performed frequently to improve accuracy, then output accuracy improves, but system complexity and computational resources increase
Solution Approach 1:
Instead of continuous re-training, the system performs partial re-training only when boundary conditions detect specific incorrect outputs, applying corrective action only when and where needed rather than excessively throughout
Solution Approach 2:
The system changes the parameter of re-training frequency from continuous to event-driven based on boundary condition monitoring, reducing computational overhead while maintaining accuracy through targeted re-training events
Data Source
AI summary
A device implemented on solid-state chips for an autonomous machine with sensors. The device includes a neural network on the autonomous machine, trained with a first training data set that includes training data generated by a sensor located remote from the autonomous machine, and configured to generate output data after processing input data. The device also includes a processor coupled to the neural network, and a detector to receive the output data and determine whether the output data breaches a predetermined condition, and a neural network manager coupled to the neural network and adapted to re-train the first neural network using another training data set if the detector determines the output data breach the first predetermined condition; and another neural network structured and trained identical to the first neural network to generate a second output data by processing the set of input data, wherein the neural networks are executed simultaneously.