DNN Bypass Paths for Autonomous Driving Perception Errors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current perception systems for automated driving face challenges in achieving high accuracy beyond level-2 ADAS, particularly in handling the long tail of rare edge cases, due to limitations in deep learning algorithms and computational resources.
Innovation Solution
An adaptable AI system that includes error resolving units capable of adapting to new errors over time, using pre-trained neural networks and adaptive steps such as adding new error resolving units, changing signatures, or adapting sensing units to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deeper and heavier neural networks are used to improve accuracy, then perception accuracy improves, but computational power requirements increase by over x10
Solution Approach 1:
The system segments the neural network into a main network and multiple error resolving units (ERPs), each specializing in specific error types. This segmentation allows the main network to remain computationally efficient while ERPs handle complex correction tasks only when errors are detected, avoiding the need for a uniformly heavy network architecture.
Solution Approach 2:
The system dynamically adjusts its computational resources by activating error resolving units only when errors are detected through the error detection mechanism. This dynamic approach allows the system to operate efficiently during normal conditions while allocating additional computational power only when and where needed to correct specific errors.
2Measurement precision
If retraining with more labeled data is performed to improve accuracy, then false-positives/false-negatives are reduced, but new false-positives/false negatives are exposed due to saturation
Solution Approach 1:
The system implements a feedback mechanism where error resolving units detect and correct errors, and this correction information feeds back to improve the system's performance. The error detection and correction loop creates continuous improvement without requiring complete retraining, allowing the system to adapt to new error types while maintaining performance on previously learned patterns.
Solution Approach 2:
Error resolving units are pre-configured with knowledge of specific error types and their corrections. When errors are detected, these pre-prepared correction mechanisms are immediately applied, avoiding the need for time-consuming retraining while still improving accuracy on previously unseen error patterns.
3Measurement precision
If the perception system is scaled to achieve 0.999 accuracy, then detection precision improves, but the system complexity and computational resources required become infeasible for in-car deployment
Solution Approach 1:
Error resolving units act as intermediaries between the main neural network and the final detection output. Instead of requiring the main network to achieve ultra-high accuracy through complex architecture, the ERPs serve as intermediary correction layers that handle specific error cases, thereby achieving high overall accuracy without proportionally increasing the complexity of the core system.
Solution Approach 2:
The system applies different levels of processing quality to different parts of the detection pipeline. The main network provides baseline detection with standard complexity, while error resolving units provide enhanced local quality correction only for specific error-prone cases. This localized enhancement achieves high overall accuracy without uniformly increasing system complexity.
Data Source
AI summary
A method that is computer implemented and is for improving an accuracy of a deep neural network (DNN) used for classification, the method includes identifying an error source within the DNN, wherein the DNN represents a deep learning model used for at least partially autonomous driving; and triggering a generation of a bypass path that bypasses the errors source.