Dynamic Vector Partitioning for Few-Shot Data Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in classifying power line components due to data scarcity and high-dimensional feature spaces, which exacerbate the curse of dimensionality and feature redundancy, hindering efficient defect detection.
Innovation Solution
A method involving dynamic vector partitioning and voting using pretrained foundation models, where feature vectors are divided into subfeature vectors and classified using distance metrics-based methods like prototypical networks, followed by voting to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models use extensive datasets for classification, then classification accuracy is improved, but data availability becomes limited in few-shot learning scenarios
Solution Approach 1:
The feature vectors are divided into multiple subfeature vectors, each representing different aspects or dimensions of the data. This segmentation allows the model to learn from limited data more effectively by focusing on specific feature subsets rather than requiring comprehensive data coverage across all features.
Solution Approach 2:
Foundation models are pre-trained on large datasets before being applied to the specific few-shot classification task. This preliminary action of pre-training enables the model to acquire general knowledge and robust feature representations that can be transferred to tasks with limited data, thereby improving classification accuracy without requiring extensive task-specific data.
2Loss of information
If foundation models extract comprehensive features from raw data, then feature representation is enriched, but feature redundancy and curse of dimensionality increase
Solution Approach 1:
The high-dimensional feature vectors extracted by foundation models are divided into multiple lower-dimensional subfeature vectors. This segmentation reduces the dimensionality of each individual feature set while preserving the overall representational capacity through the collective information from multiple subfeatures, thereby mitigating the curse of dimensionality and feature redundancy.
Solution Approach 2:
Instead of working directly with the high-dimensional feature space that causes redundancy and computational complexity, the method transforms the problem by dividing features into multiple lower-dimensional subspaces. This dimensional transformation maintains the essential information while reducing the harmful effects of high dimensionality.
Data Source
Figure 1A~1C
Figure 2
Figure 3
AI summary
The invention concerns a computer-implemented method for classifying test dataset xtest such as at least one of images, videos, audio and text. The method further comprises the steps of performing feature extraction, dividing the feature vector into a plurality of subfeature vectors, classifying each subfeature vector specific classes and performing voting to determine an optimized class of the test dataset xtest.