Training Data Distribution for Compute-Limited Driving AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated driving systems face challenges in efficiently training and utilizing machine learning models due to limited compute footprint, leading to restricted algorithm deployment and limited learning capacity, which is exacerbated by the need to handle diverse driving scenarios.
Innovation Solution
A method for determining an optimal distribution of a training dataset for machine learning models in automated driving systems, which involves selecting data samples based on a candidate distribution, training the model, evaluating its performance, and updating the distribution to achieve better model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models with deep neural networks are used to improve perception capabilities, then model performance is improved, but computational cost and resource requirements increase
Solution Approach 1:
The patent changes the parameter of training data distribution from uniform to optimized distribution based on scenario importance weights. This allows the model to achieve better performance by focusing computational resources on critical scenarios rather than uniformly processing all training data, thereby improving model performance while managing computational cost effectively
2Adaptability or versatility
If the compute footprint is increased to handle more parameters and architectural configurations, then learning capacity is improved, but hardware cost and power consumption increase
Solution Approach 1:
The patent optimizes the distribution parameter of training data to maximize learning capacity within fixed hardware constraints. By adjusting the data distribution parameters rather than increasing hardware resources, the system achieves improved adaptability and versatility without increasing device complexity or hardware cost
3Productivity
If a finite and limited training dataset is used to reduce data collection costs, then training efficiency is improved, but model performance on diverse scenarios deteriorates
Solution Approach 1:
The patent applies local quality by assigning different importance weights to different scenarios in the training data distribution. Critical scenarios receive higher weights and more focused training attention, while less critical scenarios receive lower weights. This allows the limited training dataset to be used more effectively, maintaining training efficiency while improving model reliability on diverse and critical driving scenarios
4Reliability
If more training data samples are collected to cover rare and challenging scenarios, then model robustness is improved, but data collection time and resources increase
Solution Approach 1:
The patent changes the data sampling parameter from uniform random sampling to importance-weighted sampling. This allows the system to achieve better model robustness by focusing data collection and training on rare and challenging scenarios that have higher importance weights, rather than uniformly collecting all types of data. This significantly reduces data collection time while maintaining or improving model robustness
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present inventive concept relates to a computer-implemented method (100) for determining a distribution of a training dataset for subsequent training of a machine learning model of an automated driving system, as well as other aspects thereof. The method (100) comprises: providing (S102) a first dataset by selecting, based on a candidate distribution, data samples from a second dataset of available training data; training (S104) the machine learning model on the first dataset; evaluating (S108) the machine learning model according to an evaluation criterion; and updating (S112) the candidate distribution in view of the evaluation, thereby forming an updated candidate distribution. The present inventive concept further relates to a method (200) for forming a training dataset for subsequent training of a machine learning model, as well as other aspects thereof.