Domain Adaptation for Confidence-Estimated Driving Models

Solution Overview

Problem

Traditional machine learning models are limited in handling general data and fail to perform well in unforeseen scenarios, as they are typically trained on specific data sets, leading to poor performance in recognizing unknown objects or environments encountered during autonomous driving.

Innovation Solution

A system that adapts machine learning models by training them on general data and then fine-tuning them for specific use cases through domain adaptation, using a computer system with engines for input, training, modeling, and scoring to compute accuracy scores and regressions, allowing the model to estimate unknown situations and take appropriate actions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If machine learning models are trained using data relevant to a specific use case, then the model performance for that specific function is improved, but the model fails to perform well in unforeseen scenarios or new environments

Engineering Contradiction:

Improvemodel performanceVSAvoidperformance in unforeseen scenarios

Core Design Contradiction:

ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by training machine learning models on general data that covers multiple domains and scenarios, not just specific use cases. The model is designed to handle both familiar and unforeseen situations by learning from diverse training data that includes various environments, objects, and driving conditions, making it multi-functional and adaptable to new scenarios without retraining

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent implements preliminary action by pre-training models on general data that encompasses a broad range of possible scenarios before deployment. This preliminary training on diverse data prepares the model in advance to handle unforeseen situations, allowing it to make reasonable estimates even when encountering new environments or objects during autonomous driving operations

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If machine learning models are trained on general data, then the model can handle diverse scenarios, but the model performance for specific use cases deteriorates

Engineering Contradiction:

Improveability to handle general dataVSAvoidperformance in specific functions

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent applies local quality by allowing different parts of the model or different model instances to specialize in specific functions while maintaining overall general capability. The system can apply domain adaptation techniques to fine-tune specific model components for particular use cases while preserving the general knowledge learned from diverse training data, achieving both specialization and versatility

Inventive Principle:

Principle #3Local quality

3Productivity

If machine learning models are trained on specific data sets, then the training process is efficient, but the model performance in new environments encountered during autonomous driving deteriorates

Engineering Contradiction:

Improvetraining efficiencyVSAvoidperformance in new environments

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent implements preliminary action by pre-training models on general data that encompasses a broad range of possible scenarios before deployment. This preliminary training on diverse data prepares the model in advance to handle unforeseen situations, allowing it to make reasonable estimates even when encountering new environments or objects during autonomous driving operations

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent applies universality by training machine learning models on general data that covers multiple domains and scenarios, not just specific use cases. The model is designed to handle both familiar and unforeseen situations by learning from diverse training data that includes various environments, objects, and driving conditions, making it multi-functional and adaptable to new scenarios without retraining

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS11741698B2Confidence-estimated domain adaptation for training machine learning models

Publication Date: 2023.08.29 TOYOTA JIDOSHA KK

AI summary

Embodiments of the present disclosure comprise systems and methods that implement domain adaptation. For example, some embodiments may improve standard machine learning systems by providing a best estimation of unknown situations using existing trained models. The trained models may be adapted to use in new scenarios that might not have been identified during the training phase of the machine learning. Using this adaptive approach, the model can help the vehicle system prepare a best estimate of the environment that was not identified during training and take an appropriate action.