ML Model Insight via Sensor Reconstruction for Automated Driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The black box nature of machine learning models in autonomous vehicles makes it difficult to understand how these models work and identify the causes of errors or inaccuracies in their output, hindering the improvement of advanced driver-assistance systems (ADAS) and autonomous driving (AD) systems.
Innovation Solution
The technology involves using generative AI, such as Generative Adversarial Networks or diffusion models, to reconstruct the input data as seen by the machine learning model, allowing for a comparison between the original and reconstructed data to identify potential errors or discrepancies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are used to improve the capability of automated driving systems, then the system can handle complex real-world driving scenarios, but the black box nature of these models makes it difficult to understand how they work and identify errors
Solution Approach 1:
The patent introduces an intermediary visualization system that translates the internal representations of the machine learning model into human-interpretable visual formats. This mediator layer allows users to observe what the model perceives without needing to understand the complex internal mechanics, thus resolving the contradiction between using complex models and being able to detect/understand their behavior.
Solution Approach 2:
The system creates visual copies or representations of the internal model states and sensor data processing. By generating visualizations that replicate what the model 'sees' internally, the system makes the invisible workings of the machine learning model observable, enabling error detection while maintaining the use of complex models for handling diverse driving scenarios.
2Reliability
If machine learning models are used to improve ADAS performance, then system safety and effectiveness can be enhanced, but errors in model output become harder to identify and correct
Solution Approach 1:
The patent implements a feedback mechanism where visualizations of internal model representations are continuously generated and presented to operators. This feedback loop allows operators to observe model behavior, identify errors in real-time, and understand the causes of incorrect outputs, thereby improving the ease of error correction while maintaining high system reliability through the use of advanced machine learning models.
Data Source
AI summary
The present invention relates to a computer-implemented method performed in a server. The method includes: obtaining sensor data pertaining to a scene of a surrounding environment of a vehicle equipped with an automated driving system; obtaining, for a machine learning model configured to perform a task of the automated driving system, an internal representation of the sensor data, wherein the internal representation is formed by inputting the sensor data to the machine learning model and extracting the internal representation from the machine learning model; and generating synthetic sensor data for subsequent comparison with the obtained sensor data, wherein the synthetic sensor data is generated by inputting the internal representation into a generative model trained to generate synthetic sensor data based on internal representations. The present invention further relates to a computer implemented method performed in a vehicle, as well as a server and a vehicle.