On-Board Data Selection for Autonomous Vehicle Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in efficiently selecting and processing relevant data for training machine learning models due to sparse real-world data, limited computing power at the edge, and high latency and bandwidth costs, leading to suboptimal object detection and behavior prediction.
Innovation Solution
Implementing an on-board system that computes an importance function for unlabeled data logs using a gradient-based approach, selectively transmitting data logs with high importance scores to a remote server for updating the machine learning model, thereby reducing the need for extensive data storage and processing at the edge.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If active learning is performed at the remote server to select useful data for training, then training data quality is improved, but latency and bandwidth costs increase
Solution Approach 1:
The system performs preliminary computation of the importance function at the edge device before data transmission. By calculating which data samples are most valuable for model training in advance, the system avoids the need for iterative data exchange between edge and cloud, thereby reducing latency while maintaining high training data quality selection
Solution Approach 2:
The system segments the active learning process into two parts: importance computation at the edge and model updating at the cloud. This segmentation allows the edge device to independently identify valuable data samples without requiring continuous cloud interaction, reducing bandwidth usage and latency while preserving training quality
2Loss of information
If all collected real-world data is stored and processed at the edge, then data availability is improved, but computing power and storage constraints are exceeded
Solution Approach 1:
The system extracts only the most important data samples for model training by computing an importance function at the edge. Instead of storing or processing all collected data, the system identifies and transmits only the high-value samples to the cloud, reducing edge storage and computing requirements while maintaining data availability for critical training needs
Solution Approach 2:
The system performs partial processing at the edge by computing importance scores rather than fully processing all data locally. This partial action approach allows the edge device to filter data effectively without exceeding its computing and storage constraints, while still ensuring availability of the most important data samples
3Ease of operation
If random data selection is used for training, then data collection simplicity is improved, but training effectiveness deteriorates
Solution Approach 1:
The system enables the data collection process to self-select important samples through automated importance function computation. Rather than requiring manual curation or complex selection protocols, the system autonomously identifies valuable training data based on model predictions and loss functions, maintaining operational simplicity while significantly improving training effectiveness
Solution Approach 2:
The system uses feedback from model predictions and loss function values to guide data selection. By computing the importance function based on how much a data sample would reduce model loss, the system creates a feedback loop that automatically prioritizes training-effective samples, improving reliability without complicating the data collection process
Data Source
AI summary
Systems and methods for on-board selection of data logs for training a machine learning model are provided. The system includes an autonomous vehicle having a plurality of sensors and a processor. The processor receives a plurality of unlabeled images from the plurality of sensors, a machine learning model, and a loss function corresponding to the machine learning model. For each of the plurality of images, the processor then determines one or more predictions using the machine learning model, compute an importance function based on the loss function and the one or more predictions, and transmit that image to a remote server for updating the machine learning model when a value of the importance function is greater than a threshold.