In-Vehicle Data Selection for Uncertainty-Based Model Retraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing volume of sensor data from advanced vehicle sensors for autonomous driving poses a significant computational burden, necessitating efficient data selection and transmission to maintain accurate feature detection models without excessive resource expenditure.
Innovation Solution
Implementing an in-vehicle data selection system where feature detection devices assess and transmit only sensor data with high uncertainty or low confidence metrics, allowing for targeted data transmission that improves model accuracy while minimizing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If all sensor data is transmitted to the server for model updates, then model accuracy can be improved through comprehensive data, but computational burden and resource consumption increase significantly
Solution Approach 1:
The patent extracts and transmits only the most valuable subset of sensor data - specifically data with high uncertainty or low confidence metrics - rather than transmitting all sensor data. This selective extraction reduces the data volume sent to the server while maintaining the ability to improve model accuracy through targeted training data.
Solution Approach 2:
The patent applies local quality by differentiating data based on its informational value. Instead of treating all sensor data uniformly, the system identifies and prioritizes transmission of data points that exhibit high uncertainty or low confidence, thereby allocating transmission resources to the most beneficial data while reducing overall computational burden.
2Reliability
If comprehensive sensor data is collected and transmitted, then feature detection model training can be enhanced, but data transmission and storage costs increase
Solution Approach 1:
The system extracts only the essential training data characteristics - uncertainty metrics and confidence scores - from the complete sensor data set. By transmitting these extracted features rather than raw sensor data, the patent reduces data transmission volume while preserving the information necessary for effective model training.
Solution Approach 2:
Instead of transmitting all sensor data and filtering at the server端, the patent inverts the approach by pre-filtering and selecting only the most valuable data characteristics at the vehicle端 before transmission. This reversal of the data processing sequence reduces transmission costs while maintaining training quality.
3Productivity
If selective data transmission is implemented, then resource allocation is optimized, but data selection complexity increases
Solution Approach 1:
The feature detection device performs self-service by autonomously evaluating its own confidence metrics and uncertainty levels for detected features. This self-assessment capability enables the system to automatically identify which data points warrant transmission without requiring complex external selection mechanisms, thereby optimizing resource allocation while limiting complexity growth.
Solution Approach 2:
The system incorporates feedback loops where the feature detection device continuously monitors its own performance confidence and uses this feedback to guide data selection decisions. This feedback mechanism enables adaptive data transmission that responds to actual detection quality, improving resource efficiency while using relatively simple decision logic based on confidence thresholds.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An approach is provided for selecting training observations for machine learning models. The approach involves determining a first distribution of a plurality of features observed in the training data set, and a second distribution of the plurality of features observed in the candidate pool of observations. The approach further involves selecting one or more observations in the candidate pool of observations for annotation based on the first distribution and the second distribution. The approach further involves adding the one or more observations to the training data set after annotation. The training data set is used for training the machine learning model.