Adaptable AI Signature Generation for Autonomous Driving Perception
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 edge cases, due to limitations in deep learning algorithms and computational resources.
Innovation Solution
An adaptable AI system that includes error resolving units and routers, which adapt to errors over time by generating accurate signatures to replace erroneous ones, thereby improving the accuracy of perception systems without requiring extensive retraining or increased computational power.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deeper and heavier neural networks are used to improve perception accuracy, then accuracy improves, but computational power requirements increase by over ×10
Solution Approach 1:
The system segments the neural network into a main network and multiple error resolving units (ERUs), each specialized for specific error types. This segmentation allows the main network to remain computationally efficient while ERUs handle complex correction tasks only when errors are detected, avoiding the need for a uniformly heavy network architecture.
Solution Approach 2:
Error resolving units act as intermediaries between the main neural network and the final output. When the main network produces uncertain or erroneous predictions, ERUs are invoked to correct specific error types (false positives, false negatives, classification errors), providing targeted improvement without requiring the entire system to be computationally intensive.
2Measurement precision
If retraining with more labeled data is performed to improve accuracy, then some false-positives/false-negatives are solved, but new false-positives/false negatives are exposed and system complexity increases
Solution Approach 1:
The system implements self-service through automated error detection and correction mechanisms. Error resolving units automatically identify and correct their own errors without requiring external retraining interventions. The system monitors its performance and dynamically adjusts by activating appropriate ERUs based on detected error patterns, enabling continuous improvement without external complexity.
Solution Approach 2:
The system changes parameters dynamically by adjusting which error resolving units are activated based on the detected error type and confidence levels. Instead of fixed retraining protocols, the system modifies its operational parameters in real-time, selecting appropriate correction strategies based on current performance metrics and error patterns.
3Reliability
If the neural network is made heavier to reduce error rate by factor of 2-4, then accuracy improves, but the increase in computational power makes scaling infeasible
Solution Approach 1:
The system introduces dynamics by making the network architecture adaptive rather than static. Error resolving units are dynamically activated based on the main network's confidence levels and detected error patterns. This dynamic approach allows the system to scale efficiently by only engaging additional computational resources when and where errors occur, rather than maintaining a uniformly heavy architecture throughout.
Solution Approach 2:
The system applies partial action by using error resolving units only for specific error types and confidence levels rather than applying a uniformly heavy network to all cases. The main neural network handles the majority of inference tasks with standard resources, while ERUs provide enhanced processing only when needed, achieving high reliability without excessive computational overhead across the entire system.
Data Source
AI summary
A method that is computer implemented and is for signature generation, the method includes receiving readout information, by a signature generator, the readout information is provided by a readout circuit and was extracted from a deep neural network (DNN) that was fed by a processed sensed information unit (SIU); wherein the processed SIU consists essentially of (i) SIU elements that are located within a region that is related to a bounding shape and has a desired receptive field, and (ii) padding SIU elements; and generating, by the signature generator and based on the readout information, a signature of the processed SIU for use in an at least partially autonomous driving of a vehicle.