Image Classification with Continual Learning Against Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Catastrophic forgetting occurs in convolutional neural networks (CNNs) when new tasks are learned, leading to a decrease in precision for old tasks.
Innovation Solution
An image processing technology that combines continual learning with metric learning, using two neural networks to maintain the basic class knowledge while updating an additional neural network for new classes, with centroid vector correction to minimize forgetting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CNN parameters are re-trained 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 divides the neural network into two separate networks: a basic neural network that maintains knowledge of base classes and an additional neural network that learns new classes. This segmentation prevents the basic network from forgetting old tasks while the additional network adapts to new tasks, resolving the contradiction between adaptability and precision for old tasks.
Solution Approach 2:
The patent introduces a selection unit that acts as an intermediary to determine whether to use the basic neural network or the additional neural network based on the input data. This mediator selects the appropriate network for classification, ensuring that old tasks are handled by the preserved basic network while new tasks are handled by the adapted additional network.
2Adaptability or versatility
If continual learning is performed by training the model from scratch on new tasks, then new tasks are learned effectively, but time and computational resources are lost due to re-training
Solution Approach 1:
The patent pre-trains a basic neural network on base classes before encountering new tasks. This preliminary action creates a foundation that can be quickly adapted to new tasks through the additional neural network, avoiding the need to re-train from scratch and significantly reducing training time for new tasks.
Solution Approach 2:
The patent implements a dynamic system where the selection unit determines which neural network to use based on the input data characteristics. The system dynamically switches between the basic neural network for base classes and the additional neural network for new classes, enabling efficient adaptation without full re-training.
3Device complexity
If a single neural network is used for both base classes and new classes, then device complexity is reduced, but catastrophic forgetting occurs causing performance degradation
Solution Approach 1:
The patent segments the classification system into two distinct neural networks: a basic neural network specialized for base classes and an additional neural network for new classes. This segmentation prevents catastrophic forgetting by isolating the knowledge of base classes in the basic network while allowing the additional network to learn new classes without interference.
Solution Approach 2:
The patent creates an additional neural network that copies the structure and initial knowledge from the basic neural network, then adapts it for new classes. This copying approach allows the system to maintain the reliable basic network while creating a specialized additional network for new tasks, preserving overall classification accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A basic class selection unit (20) selects, in response to input data, a base class based on an embedding vector output by a basic neural network that has learned the base class and a centroid vector of the base class. A continual learning unit (50) continually learns an additional class by using an additional neural network that has learned the base class. An additional class selection unit (40) selects, in response to the input data, an additional class based on an embedding vector output by the additional neural network subjected to continual learning and centroid vectors of the base class and the additional class. A classification determination unit (80) classifies the input data based on the base class selected by the base class selection unit (20) and the additional class selected by the additional class selection unit (40).