Ambiguous Zone Embedding Routing for Low-Confidence Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a growing need to increase the accuracy of classification in neural networks used in vehicles, particularly when the initial classification process is inaccurate.
Innovation Solution
A method for self-supervised learning of an ambiguous zone in an embedding space involves identifying embeddings with low classification confidence levels, training additional neural networks in a self-supervised manner, and associating these embeddings with routing rules for accurate classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a first neural network is trained using supervised learning, then it can classify sensed information units, but it produces inaccurate classifications for certain items falling into ambiguous zones
Solution Approach 1:
The embedding space is segmented into distinct regions: confident zones and ambiguous zones. The system identifies and separates embeddings with low classification confidence levels from those with high confidence, allowing different processing strategies for each region. This segmentation enables targeted improvement of classification accuracy specifically for ambiguous cases without reprocessing all data.
Solution Approach 2:
A second neural network is introduced as an intermediary component to handle ambiguous cases. This second network receives embeddings from the ambiguous zone and performs additional classification analysis, acting as a mediator between the first neural network's uncertain outputs and the final classification decision, thereby improving overall reliability.
2Reliability
If additional neural networks are trained to handle ambiguous zones, then classification accuracy improves, but system complexity increases
Solution Approach 1:
Instead of improving the entire neural network system uniformly, the solution applies enhanced processing only locally to ambiguous zones. The second neural network is specifically deployed for embeddings with low classification confidence levels, while high-confidence classifications continue to be handled by the first network alone. This localized approach improves accuracy where needed while minimizing overall system complexity.
Solution Approach 2:
The system uses self-supervised learning on embeddings from ambiguous zones, allowing the second neural network to learn from the data itself without requiring additional labeled training data. This self-service mechanism reduces the complexity of the training process and eliminates the need for manual annotation of ambiguous cases.
3Measurement precision
If self-supervised learning is applied to ambiguous zones, then classification precision improves, but training time and computational resources increase
Solution Approach 1:
The system applies self-supervised learning only partially, specifically to embeddings identified as belonging to ambiguous zones rather than training on the entire dataset. By focusing computational resources only on the subset of data that requires improvement, the training time and resource consumption are significantly reduced while still achieving enhanced classification precision for problematic cases.
Data Source
AI summary
A method for self-supervised learning of ambiguous zone embedding space, the method includes identifying, by a processing circuit and during a validation process of a first neural network, a set of embeddings that represent a group sensed information units that are associated with a classification confidence level below a threshold; wherein the first neural network was trained by a supervised training process; the set of embeddings defining an ambiguous zone; and triggering a training of a second neural network, in a self-supervised learning process, across the group of sensed information unit.