Structure-Based Training: A Training Method Aimed at Pixel Errors for a Correlation-Coefficient-Based Neural Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• One-shot learning neural networks without pre-training face performance limitations due to insufficient information from single training samples.
• Existing networks typically optimize parameters via back-propagation, which can be suboptimal for certain architectures.

Purpose of the Study:

• To improve the performance of correlation-coefficient-based neural networks in one-shot learning scenarios.
• To develop a novel training strategy that leverages the full potential of the support set and addresses network structural limitations.

Main Methods:

• Design of three types of 'shadow nodes' to enhance network adaptability.
• Proposal of a structure-based training method that optimizes the network's architecture by correcting pixel errors, rather than relying solely on back-propagation.
• Focus on training strategies for branches that are either unexpectedly activated or inactivated.

Main Results:

• Significant improvements in Intersection over Union (IOU) for target detection across multiple datasets: 4.83% on Fashion-Mnist, 4.02% on Omniglot, and 3.89% on Cifar-10.
• A notable decrease of 27.32% in misclassifications of category '7' as '1' on the Mnist dataset after training.
• Demonstrated enhancement of the correlation-coefficient-based network's practicality and ability to learn from accumulating reliable samples.

Conclusions:

• The proposed shadow nodes and structure-based training method enhance correlation-coefficient-based networks, making them more practical for target detection, especially in scenarios with limited initial data.
• This approach offers a non-gradient-based parameter optimization strategy and mimics human-like learning from few references, advancing one-shot learning research.