Classifier Training Using Semantic Similarity for New Categories
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks face inefficiencies in re-training when adapting to new categories due to the need to re-learn features, leading to increased workload and performance degradation.
Innovation Solution
Utilizing semantic similarity to identify target samples and adjust training based on multi-view features, incorporating a method that includes determining semantic transfer loss and balanced learning loss functions to enhance training efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the feature extractor re-learns features for new categories using novel samples, then the classifier can adapt to new categories, but the training workload increases
Solution Approach 1:
The patent introduces semantic similarity as an intermediary mechanism to bridge base samples and novel samples. By calculating semantic similarity between novel samples and base samples, the system can identify which base samples are most relevant for transferring knowledge to new categories, thereby reducing the need for the feature extractor to re-learn all features from scratch while maintaining adaptability to new categories
Solution Approach 2:
The patent changes the training parameter by incorporating semantic similarity scores as a weighting factor. Instead of treating all base samples equally, the system adjusts the influence of each base sample based on its semantic similarity to the novel sample, allowing the classifier to focus on the most relevant knowledge transfer and reduce overall training workload
2Reliability
If more base samples are used for knowledge transfer, then the classifier performance improves, but the training time increases
Solution Approach 1:
The patent applies local quality by differentiating the treatment of base samples based on their semantic similarity to novel samples. Rather than uniformly processing all base samples, the system identifies and prioritizes base samples with high semantic similarity to the novel category, allowing the classifier to achieve good performance by focusing computational resources on the most relevant local subset of base samples
Data Source
AI summary
This application relates to Artificial intelligence and provides a method for training a classifier, one example method including: obtaining a first training sample, where the first training sample includes a corresponding semantic tag; obtaining a plurality of second training samples, where each of the second training samples includes a corresponding semantic tag; determining a target sample from the plurality of second training samples based on semantic similarities between the first training sample and the plurality of second training samples; and training the classifier based on the first training sample, the target sample, and a semantic similarity between the first training sample and the target sample.