Neural Network Correction Unit for Bias and Weight Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks primarily correct bias parameters, limiting their adaptability to recognition tasks, as they do not adjust weight parameters across different classes, restricting their performance in pattern recognition.
Innovation Solution
An arithmetic operation apparatus that corrects both bias and weight parameters of neural network units across multiple classes by incorporating a correction unit layer, allowing for more comprehensive feature adaptation and recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional neural networks only correct bias parameters, then the device complexity is low, but the adaptability to recognition tasks is limited
Solution Approach 1:
The patent segments the correction function into two distinct components: bias parameter correction and weight parameter correction. By dividing the correction mechanism into separate functional units (bias correction unit and weight correction unit), the system can independently adjust each parameter type, thereby improving adaptability without creating an intractably complex system. This segmentation allows for modular implementation and targeted optimization of specific correction functions.
Solution Approach 2:
The patent extends the correction mechanism from a single dimension (bias parameters only) to multiple dimensions by incorporating weight parameter correction. This dimensional expansion allows the neural network to adapt along additional parameter axes, significantly enhancing recognition task adaptability. The multi-dimensional correction approach enables the system to address both inter-class and intra-class variations simultaneously.
2Measurement precision
If neural networks use predetermined classes only, then the device complexity is low, but the recognition performance for beyond-predetermined classes is poor
Solution Approach 1:
The patent implements dynamic class adaptation by enabling the neural network to correct weight parameters specific to each class. This dynamic adjustment allows the system to adapt its recognition behavior based on the specific class being evaluated, rather than using fixed predetermined classes. The dynamic weight correction mechanism enables the network to optimize its recognition performance for each class individually, improving overall measurement precision.
Solution Approach 2:
The patent applies parameter changes by systematically adjusting both bias and weight parameters based on recognition results. By changing the weight parameters dynamically according to class-specific performance, the system can optimize recognition accuracy for beyond-predetermined classes. This parameter adaptation mechanism allows the neural network to evolve its recognition capabilities beyond static class definitions.
3Adaptability or versatility
If neural networks correct both bias and weight parameters, then the adaptability improves, but the calculation complexity increases
Solution Approach 1:
The patent segments the correction process into distinct stages: bias correction, weight correction, and recognition. By dividing the calculation into separate functional steps, the system can manage calculation complexity more effectively. Each segment handles a specific aspect of parameter correction, allowing for optimized computation at each stage rather than attempting to correct all parameters simultaneously.
Solution Approach 2:
The patent applies local quality by implementing class-specific weight correction rather than global uniform correction. Each class receives tailored weight adjustments based on its specific recognition characteristics, allowing for more efficient calculation focused on relevant parameters. This localized correction approach reduces unnecessary calculations while maintaining high adaptability for feature adaptation.
Data Source
AI summary
According to an embodiment, an arithmetic operation apparatus for a neural network includes an input layer calculator, a correction unit calculator, a hidden layer calculator, and an output layer calculator. The input layer calculator is configured to convert an input pattern into features as outputs of an input layer. The correction unit calculator is configured to perform calculation on N unit groups corresponding respectively to N classes of the input pattern and including correction units that each multiply a value based on inputs by a weight determined for the corresponding class. The hidden layer calculator is configured to perform calculation in a hidden layer based on the outputs of the input layer, another hidden layer, or the correction unit calculator. The output layer calculator is configured to perform calculation in an output layer based on the calculation for the hidden layer or the outputs of the correction unit calculator.