Machine learning device, machine learning method, and machine learning program
By optimizing feature vectors and classification weights using a graph neural network or graph attention network, the machine learning device addresses the separation issue in incremental learning, enhancing classification accuracy.
JP7803120B2Active Publication Date: 2026-01-21JVC KENWOOD CORP
Patent Information
- Application Number
- JP2021209555
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-12-23
- Publication Date
- 2026-01-21
- Estimated Expiration
- 2041-12-23
Smart Images
Figure 0007803120000001 
Figure 0007803120000002 
Figure 0007803120000003
Abstract
To provide a machine learning technique for improving class classification accuracy.SOLUTION: A pre-trained feature extraction unit 54 extracts, by using a pre-trained model, feature vectors of samples in a base class. A base class classification weight 58 is for receiving an input of feature vectors of the samples in the base class, to classify the samples in the base class by using the classification weight of the base class. A feature optimization unit 64 performs meta-learning of an optimization module by using the pre-trained model as a basis, and optimizes the feature vectors of samples in new classes. A new class feature averaging unit 66 averages, for each of the new classes, the feature vectors of the samples in the new class, and calculates the classification weight of the new class. A graph neural network 70 receives input of the classification weight of the base class and classification weights of the new classes, performs meta-learning of dependence relationships between the base class and the new classes, and outputs reconstruction classification weights.SELECTED DRAWING: Figure 6
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Feature offset correction method based on meta-learning
CN112116063A