Interactive neural network training in adverse conditions for revoking bias in driving
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Autonomous vehicles face challenges in accurately classifying road elements due to bias in sensing data, particularly under adverse conditions such as partial occlusions and varying illumination, which affects the reliability of decision-making processes.
Innovation Solution
A method involving the generation of artificially generated sensed information units to create a diverse dataset that includes biased and non-biased data, used to train machine learning processes to treat each road element separately, thereby overcoming classification bias and enhancing the ability to identify road elements under various conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning processes are trained on real sensed information units from adverse conditions, then the system learns to handle real-world variability, but classification bias occurs when multiple road elements appear together
Solution Approach 1:
The patent creates artificial copies of sensed information units by generating synthetic road elements that can be inserted into existing images. These synthetic copies replicate the visual appearance and characteristics of real road elements while allowing controlled manipulation of their presence and combination, enabling the system to learn without acquiring actual biased data
Solution Approach 2:
The system manipulates parameters of the sensed information units by adjusting which road elements are present, their positions, and their combinations. By systematically varying these parameters to create different scenarios including biased combinations and unbiased individual elements, the training process learns to distinguish true classifications from bias-induced errors
2Loss of information
If the system processes all detected road elements together, then comprehensive scene understanding is achieved, but bias causes incorrect classification of individual elements
Solution Approach 1:
The patent segments the classification process by treating each road element as a separate classification task rather than a single combined classification. The system generates and processes synthetic examples where individual elements appear alone as well as in combination, enabling the model to learn element-specific features independent of contextual bias while maintaining comprehensive scene understanding
3Adaptability or versatility
If diverse training data is used to improve generalization, then robustness to varying conditions is enhanced, but classification bias persists under adverse conditions
Solution Approach 1:
The system performs preliminary action by pre-generating a comprehensive set of synthetic road elements and biased/unbiased combinations before the actual classification task. This preparatory step creates a balanced training dataset that proactively addresses potential bias scenarios, allowing the model to learn correct classification patterns before encountering real biased data during deployment
Data Source
AI summary
A method of interactive neural network training for driving, the method includes identifying, across a first set of images of road elements and using a neural network to output first driving related outcomes, an image comprising a combination of elements in an initial scenario that is below a confidence level threshold; determining the combination in the initial scenario as a bias; interacting, responsive to the determining, with a second set of images, using the neural network to output second driving related outcomes, wherein the second set of images are created, at least in part, artificially; and revoking, with the second process running interactively with the first process, the determined bias in the first process, by interacting with the first process using the second driving related outcomes of the second process.