Neural Network Learning Process Evaluation via Layer Statistics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting physical and chemical phenomena, such as chemical reactions, lack effective evaluation support for the learning process of prediction models, particularly in optimizing reactor systems.
Innovation Solution
A method and device for supporting the evaluation of a prediction model's learning process by training a neural network with an input layer, intermediate layer, and output layer, and outputting statistical information on input values to these layers, including frequency distributions and activation functions, to determine peak validity within predetermined ranges.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If statistical information on input values to intermediate and output layers is output, then evaluation support technology is improved, but device complexity increases
Solution Approach 1:
The patent extracts and outputs statistical information (frequency distributions, peak values) from the neural network's intermediate and output layers for evaluation purposes. This extraction enables objective evaluation of the learning process without requiring complete transparency of the entire model, thus improving evaluation support while managing complexity through selective information extraction.
Solution Approach 2:
The patent segments the evaluation process into distinct components: statistical information extraction from intermediate layers, statistical information extraction from output layers, and separate evaluation determinations. This segmentation allows the complex evaluation task to be broken down into manageable parts, improving reliability while controlling overall system complexity.
2Reliability
If frequency distributions and peak values are determined and displayed, then learning process evaluation is improved, but information processing time increases
Solution Approach 1:
The patent performs preliminary computation of statistical information (frequency distributions, peak values) during or after the training process, so that when evaluation is needed, the data is already prepared. This preliminary action reduces the time required for actual evaluation while maintaining objective and reliable assessment of the learning process.
Solution Approach 2:
The patent computes and displays only the essential statistical characteristics (frequency distributions and peak values) rather than all possible model parameters. This partial action provides sufficient information for reliable evaluation while significantly reducing processing time and computational overhead.
3Measurement precision
If initial value of weight coefficient is determined based on predetermined ranges, then prediction accuracy is improved, but device complexity increases
Solution Approach 1:
The patent determines initial values of weight coefficients by changing parameters within predetermined ranges (0.01-0.99 for output values, or differential values >0 for activation functions). This parameter-based approach enables objective evaluation and optimization of prediction accuracy while maintaining manageable complexity through standardized parameter selection criteria.
Data Source
AI summary
A method is executed by an information processing device for supporting evaluation of a learning process of a prediction model. The method includes: training a neural network model that includes an input layer, an intermediate layer, and an output layer, based on actual data including an explanatory factor and an objective factor; and outputting statistical information on input values that are input to the intermediate layer and the output layer of the neural network model.