Neural Network Hidden Layer Role Analysis via Unified Community Structure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies cannot effectively analyze the similarity of roles between layers in a multilayer neural network, limiting the understanding of the input/output mapping performed in the entire hidden layer.
Innovation Solution
An analysis device and method that calculates the strength of relationships between input data dimensions, neural network units, and output data dimensions, using nonnegative matrix factorization to determine the main roles and associations within the hidden layer, enabling the analysis of input/output mapping across the entire neural network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If clustering of units is performed for each layer of the neural network, then the role of each community in a community structure can be extracted, but it is not possible to grasp the similarity of the role of each unit between layers
Solution Approach 1:
The patent merges the analysis across multiple layers by constructing a unified community structure that incorporates units from all layers. Instead of performing clustering independently for each layer, the invention combines the adjacency matrices of all layers to create a comprehensive community structure, enabling both precise role extraction and cross-layer comparison.
Solution Approach 2:
The patent creates a universal analysis framework that can handle both layer-specific role extraction and cross-layer similarity assessment. The community structure extraction method is designed to be multi-functional, serving both to identify local roles within layers and to enable global role comparison across layers through the unified structure.
2Measurement precision
If existing community structure extraction methods are used, then the role of each community can be determined, but it is not possible to analyze what type of input/output mapping is mainly performed in the entire hidden layer
Solution Approach 1:
The patent extends the analysis from the traditional single-layer dimension to a multi-layer dimensional space. By constructing a unified community structure that spans multiple layers, the invention enables analysis in this extended dimension, allowing simultaneous determination of community roles and identification of input/output mapping patterns across the entire hidden layer.
Solution Approach 2:
The patent segments the analysis into two complementary components: community structure extraction for role determination, and input-output mapping analysis for functional understanding. This segmentation allows the system to maintain precise role determination while simultaneously enabling comprehensive analysis of the entire hidden layer's mapping behavior.
Data Source
AI summary
It is possible to analyze what type of input/output mapping is mainly performed in the entire hidden layer of a neural network. A relationship analysis unit 30 calculates strength of a relationship of each combination of a dimension of the input data and a unit of the neural network and calculates strength of a relationship of each combination of the unit and a dimension of the output data. A role analysis unit 32 calculates a relationship between a prescribed number of types of roles and the unit and a relationship between the prescribed number of types of roles, the dimension of the input data, and the dimension of the output data on the basis of the strength of the relationship of each combination of the dimension of the input data and the unit of the neural network and the strength of the relationship of each combination of the unit and the dimension of the output data.