Image Classification Using Averaged Feature Matrices for Incremental Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification technologies using deep neural networks face challenges in maintaining accuracy when adapting to new tasks without forgetting old tasks, leading to catastrophic forgetting, and struggle with overfitting when trained on small datasets.
Innovation Solution
An image classification apparatus and method that employs a feature extraction unit to generate deep and shallow feature vectors, calculates averaged feature matrices, and replaces weight matrices with these matrices to improve classification accuracy in incremental learning scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If deep neural network parameters are re-learned in response to the entirety of the dataset to adapt to new tasks, then adaptability to new tasks is improved, but precision for old tasks deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent segments the feature extraction process into multiple independent feature extractors, each specialized for specific tasks or data distributions. This allows the system to maintain separate feature representations for different tasks, preventing catastrophic forgetting while enabling adaptation to new tasks through the segmented architecture.
Solution Approach 2:
The patent introduces a new dimension by computing feature vectors from multiple different feature extractors simultaneously. Instead of relying on a single feature space, the system operates in a multi-dimensional feature space where each dimension corresponds to a different extractor's output, allowing preservation of old task knowledge while adapting to new tasks.
2Ease of operation
If deep learning models are trained on a small number of images, then ease of operation is improved, but manufacturing precision deteriorates due to overfitting
Solution Approach 1:
The patent merges feature vectors from multiple different feature extractors to create a comprehensive feature representation. By combining features from multiple sources trained on different data distributions, the system achieves better generalization performance even when individual extractors are trained on small datasets, reducing overfitting through feature fusion.
Solution Approach 2:
The patent creates feature extractors with multi-functionality that can handle both base tasks and incremental tasks. The averaged feature matrix serves as a universal representation that works across different tasks and data distributions, allowing the system to maintain good performance with limited data per task while avoiding overfitting.
Data Source
AI summary
A feature extraction unit outputs first and second feature vectors of an input image. An averaged first/second feature calculation unit calculates an averaged first/second feature vector by averaging first/second feature vectors of a given class and obtains an averaged first/second feature matrix by aggregating averaged first/second feature vectors of all classes. A first/second feature similarity calculation unit calculates a first/second similarity from the first/second feature vector of the input image and a first/second weight matrix. The averaged first/second feature calculation unit replaces the first/second weight matrix of the first/second feature similarity calculation unit with the averaged first/second feature matrix.