一种单权重双向预测编码方法及神经网络训练方法
By introducing a single-weight bidirectional predictive coding method into deep neural networks and using both forward and backward errors to update the state of intermediate layer nodes, the biological rationality and robustness issues in deep network training are solved, achieving high-precision learning on real datasets and enhanced robustness in noisy environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA NANHU ACAD OF ELECTRONICS & INFORMATION TECH
- Filing Date
- 2023-03-15
- Publication Date
- 2026-07-17
AI Technical Summary
Existing deep neural networks lack biological rationality and robustness during training, and are particularly prone to collapse and sensitive to noise in deep networks.
A single-weight bidirectional predictive coding method is adopted. By defining forward and backward function chains, the parameters and node states of the intermediate layer are shared, and the node states of the intermediate layer are updated based on the common gradient descent of the forward and backward errors. The feedback link is added to enhance the biological rationality and robustness.
It achieves good learning performance and high accuracy of deep neural networks on real datasets, improves the robustness of the model, and enables it to learn multiple times in noisy environments.
Smart Images
Figure CN118673981B_ABST