Incremental Embedding Clustering 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
An iterative incremental learning method that identifies items with below-threshold classification confidence and creates additional clusters, improving classification accuracy over time by increasing the number of iterations and imposing stricter detection thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the classification threshold is increased to improve accuracy, then classification precision improves, but fewer items are classified including rare items
Solution Approach 1:
The patent implements dynamic threshold adjustment where the classification threshold is not fixed but adapts over time through iterative learning. The system starts with a lower threshold to capture more items including rare ones, then progressively increases the threshold as it learns from misclassifications, achieving both broad coverage and high precision at different stages
Solution Approach 2:
The system performs self-learning by automatically identifying its own misclassifications and using them to improve future classifications. Through iterative incremental learning, the system serves itself by continuously refining its classification capabilities without external intervention, progressively improving accuracy while maintaining coverage
2Measurement precision
If iterative incremental learning is applied to improve classification of rare items, then classification accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by focusing iterative learning only on misclassified items rather than reprocessing all items in each iteration. The system identifies and learns from errors selectively, achieving improvement in classification accuracy without the computational cost of complete reprocessing at each step
Solution Approach 2:
The system performs preliminary classification with an initial threshold to quickly process the majority of items. Only items that fail this preliminary classification are subjected to further iterative learning, allowing most items to be processed rapidly while still achieving high accuracy for difficult cases
Data Source
AI summary
A method for incremental learning classification capabilities, the method includes identifying, at each iteration of an iterative incremental learning process, that an embedding exhibits a clustering confidence level that is below a threshold, the embedding, generated at least in part by a machine learning process, representing a detector output that is responsive to a sensed information unit; identifying, at each iteration by accessing a data structure associated with one or more reference detector output, signatures that are similar to a signature of the detector output; and determining, at each iteration, an additional cluster for an embedding associated with the detector output and for reference embeddings associated with the one or more reference detector output signatures. Such that at each iteration of the iterative incremental process the identifying is based on at least one more determined cluster than a preceding iteration, and road elements associated with embeddings that fall within the determined cluster exhibit a clustering confidence level that is above the threshold and are classified, during inference, in accordance with the determined cluster.