Neural Network Ensemble for Robust Pattern Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks struggle to capture minute pattern variations essential for recognition while maintaining robustness against unrelated pattern variations, such as image displacements, which are lost due to pooling layers used for robustness.
Innovation Solution
An information processing device and method that perform multiple preprocesses on patterns to create different degrees of pattern variation, determining a network structure for each preprocess to learn a neural network with associated robustness, allowing for the integration of neural networks to capture minute variations while maintaining robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a pooling layer is introduced to improve robustness against positional variation, then robustness against displacement is improved, but minute characteristic differences are lost
Solution Approach 1:
The patent divides the neural network into multiple specialized networks, each responsible for processing patterns with specific types or degrees of variation. Instead of using a single pooling layer that loses information, the system segments the processing task across multiple networks that can be selectively applied based on the input pattern's characteristics.
Solution Approach 2:
The patent introduces dynamic selection of neural network architectures based on the input pattern's variation characteristics. The system dynamically determines which network to use by analyzing the pattern's properties, allowing the robustness against displacement to be adjusted dynamically rather than being fixed by a static pooling layer structure.
2Reliability
If deep learning is used to achieve robust pattern recognition, then robustness against pattern variation is improved, but ability to capture minute pattern variations deteriorates
Solution Approach 1:
The patent applies different processing characteristics to different parts of the pattern recognition system. Each neural network is specialized for specific types of patterns or variation degrees, creating local quality differences in the system's response to different input characteristics. This allows minute variations to be preserved in specialized networks while other networks handle robustness.
Solution Approach 2:
The patent changes the parameters and architecture of neural networks based on the input pattern's characteristics. By analyzing the degree and type of pattern variation, the system adjusts which network parameters and architectures are applied, enabling adaptive balance between robustness and precision for different input scenarios.
Data Source
AI summary
A neural network capturing a minute pattern variation useful for recognition while maintaining robustness against a pattern variation unrelated to recognition is learned. A preprocessing unit performs, on a set of patterns being to be learned and including a specific pattern variation, a plurality of preprocesses causing different degrees of the specific pattern variation. A network structure determination unit determines, for each of the plurality of preprocesses, a network structure of a neural network having robustness according to a degree of the specific pattern variation after the preprocess. A network learning unit learns, for each of the plurality of preprocesses, the neural network with the network structure associated with the preprocess using the set of patterns after the preprocess.