On-Device ML Network Learning for Faster Autonomous Vehicle Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing on-device learning methods for autonomous vehicles are hindered by slow update cycles and limited accessibility due to reliance on cloud servers and Over-The-Air (OTA) connections, which are restricted in certain environments.

Innovation Solution

The method employs multi-stage learning with adaptive hyper-parameter sets, allowing the on-device learning device to divide the learning process into stages, generate adaptive hyper-parameter sets, and train the machine learning network using both previous and new training data without cloud connectivity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If cloud-based training with OTA connection is used, then machine learning network can be updated with new data, but update cycle becomes slow and accessibility is limited

Engineering Contradiction:
Improveadaptability to new driving environmentVSAvoidupdate cycle time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The training process is segmented into multiple stages (first stage training, second stage training, etc.) where each stage uses different hyperparameter sets and different proportions of new versus previous training data. This allows incremental adaptation without requiring complete retraining or cloud connectivity, reducing update time while maintaining adaptability.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The hyperparameter sets are dynamically adjusted across different training stages. The learning rate, batch size, and other hyperparameters change adaptively based on the training stage and the proportion of new training data, enabling efficient on-device learning without fixed configuration constraints.

Inventive Principle:
Principle #15Dynamics

2Productivity

If multi-stage learning with adaptive hyperparameter sets is used, then learning speed improves, but system complexity increases

Engineering Contradiction:
Improvelearning speedVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

Multiple hyperparameter sets are prepared in advance for different training stages. The first hyperparameter set is used for initial training with a higher proportion of new data, while subsequent hyperparameter sets are used for fine-tuning with varying data proportions. This preliminary preparation enables faster learning without requiring complex real-time optimization during training.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If on-device training with limited computing power is used, then cloud connectivity dependency is reduced, but training efficiency decreases

Engineering Contradiction:
Improveindependence from cloud connectivityVSAvoidtraining efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The system changes key training parameters including the proportion of new versus previous training data, learning rates, batch sizes, and number of training epochs across different stages. These parameter adjustments optimize training efficiency for on-device computation with limited power, achieving effective adaptation without cloud dependency while maintaining reasonable training speed.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3910563B1Method for performing on-device learning of machine learning network on autonomous vehicle by using multi-stage learning with adaptive hyper-parameter sets and device using the same
Publication Date: 2025.05.21 STRADVISION
  • EP3910563B1 patent drawingFigure 1
  • EP3910563B1 patent drawingFigure 2
  • EP3910563B1 patent drawingFigure 3

AI summary

A method for performing on-device learning of embedded machine learning network of autonomous vehicle by using multi-stage learning with adaptive hyper-parameter sets is provided. The processes include: (a) dividing the current learning into a 1-st stage learning to an n-th stage learning, assigning 1-st stage training data to n-th stage training data, generating a 1_1-st hyper-parameter set candidate to a 1_h-th hyper-parameter set candidate, training the embedded machine learning network in the 1-st stage learning, and determining a 1-st adaptive hyper-parameter set; (b) generating a k_1-st hyper-parameter set candidate to a k_h-th hyper-parameter set candidate, training the (k-1)-th stage-completed machine learning network in the k-th stage learning, and determining a k-th adaptive hyper-parameter set; and (c) generating an n-th adaptive hyper-parameter set, and executing the n-th stage learning, to thereby complete the current learning.