Vehicle Classification Error Correction Without Neural Network Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks in vehicles often output erroneous classification decisions, requiring time-consuming and costly software updates for retraining to correct these errors.
Innovation Solution
A system comprising a drive related unit, analyze unit, and fix unit that automatically identifies and corrects erroneous classification decisions without retraining the neural network, using auto-labeling capabilities and computational resources to generate automated ground truth labeling and signatures for correcting faulty neural network outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network retraining is performed to correct erroneous classification decisions, then classification accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The patent introduces an intermediary correction system that sits between the neural network and the classification output. This system includes a correction module that receives erroneous classifications, generates corrected predictions using a correction model, and outputs the corrected results. This intermediary approach allows error correction without modifying the original neural network weights or requiring retraining time.
Solution Approach 2:
The patent creates a correction model that copies or replicates the behavior of the original neural network for specific error patterns. Instead of retraining the entire neural network, a separate correction model is trained on erroneous predictions and their correct counterparts, then used to correct future errors. This copying approach enables efficient error correction without affecting the original network.
2Measurement precision
If neural network retraining is performed to correct erroneous classification decisions, then classification accuracy is improved, but software update complexity and cost increase
Solution Approach 1:
The patent segments the error correction function from the main neural network into a separate correction module. This segmentation allows the correction model to be independently trained, validated, and deployed without affecting the original neural network. The correction module can be updated separately and applied to correct errors without requiring complex full-system software updates.
Solution Approach 2:
The correction module acts as an intermediary layer that can be independently updated and deployed. This intermediary structure simplifies software updates because only the correction model needs to be updated, not the entire neural network system. The correction module receives input from the original network and outputs corrected predictions, maintaining system modularity.
3Measurement precision
If neural network state is changed to correct errors, then classification accuracy improves, but system stability and existing performance degrade
Solution Approach 1:
The correction module serves as an intermediary that corrects errors without modifying the original neural network state. The original network weights and architecture remain unchanged, preserving system stability and existing performance characteristics. The correction model operates separately and only affects the final output predictions, not the underlying network structure.
Solution Approach 2:
The patent applies partial correction action by only modifying the specific erroneous predictions rather than changing the entire neural network state. The correction model is applied selectively to correct classified decisions that fall within its training scope, leaving the rest of the neural network unchanged and stable.
Data Source
Figure 1(A)~1(D)
Figure 2
Figure 3
AI summary
A method for real time management of detected issues, the method includes producing, by a classification unit having a neural network, a classification decision for sensed information obtained in an environment of a vehicle; generating, by one or more computing devices with auto-labeling capabilities, an automated ground truth labeling for the sensed information; detecting, by the one or more computing devices and based on a performance indication related to the automated ground truth labeling, an issue with respect to the classification decision; and responsive to the detecting, addressing the detected issue in a driving in the environment of the vehicle by a computer device associated with the vehicle, using a signature generated in association with at least the classification decision or with the detected issue. The neural network is in a same state in the producing of the classification decision, the detecting the issue, and the addressing the detected issue.