Hierarchical Neural Network for Multi-Condition Data Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning techniques struggle to effectively learn data sets that belong to the same category but are acquired under different conditions, such as images captured using different devices, at varying times, or with different exposures, as they do not adequately account for these conditions during the learning process.
Innovation Solution
A learning apparatus and method utilizing a hierarchical network with independent input layers for data from different conditions, sharing an intermediate layer for feature quantity calculation, which reduces the risk of overlearning and allows for appropriate learning of data from the same category acquired under different conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If data from the same category acquired under different conditions are input to a single input layer, then the network structure is simplified, but the learning accuracy deteriorates because the different acquisition conditions are not adequately accounted for
Solution Approach 1:
The patent divides the single input layer into multiple independent input layers, with each input layer dedicated to processing data from a specific acquisition condition. This segmentation allows the network to independently process and learn features from different acquisition conditions without interference, thereby maintaining learning accuracy while managing model complexity through modular architecture.
Solution Approach 2:
The patent introduces an additional dimension to the network architecture by adding independent input layers for different acquisition conditions. This dimensional expansion allows the network to simultaneously process multiple data sources with different characteristics, enabling better representation of acquisition condition variations without collapsing the structural complexity.
2Measurement precision
If multiple independent input layers are added to process data from different acquisition conditions, then the learning accuracy improves, but the device complexity increases
Solution Approach 1:
The patent merges multiple independent input layers with a shared intermediate layer, creating a hierarchical structure where feature extraction is performed separately for each acquisition condition but then combined through the shared intermediate layer. This merging approach allows the network to capture condition-specific features while maintaining a unified representation, improving learning accuracy without proportionally increasing overall model complexity.
Solution Approach 2:
The shared intermediate layer serves multiple functions simultaneously: it processes features from different acquisition conditions, extracts common representations, and reduces the impact of condition-specific variations. This multi-functionality allows the network to achieve high learning accuracy while avoiding the need for completely separate processing paths for each condition, thereby controlling device complexity.
3Reliability
If a hierarchical network with shared intermediate layer is used, then the risk of overlearning is reduced, but the calculation process becomes more complex
Solution Approach 1:
The shared intermediate layer acts as an intermediary between the multiple independent input layers and the subsequent processing layers. It mediates the feature extraction process by receiving and processing features from different acquisition conditions, thereby preventing overlearning through a unified representation while organizing the calculation process into manageable stages: input processing, intermediate feature extraction, and final output generation.
Data Source
AI summary
An object of the present invention is to provide a learning apparatus and a learning method capable of appropriately learning pieces of data that belong to the same category and are acquired under different conditions. In a learning apparatus according a first aspect of the present invention, first data and second data are respectively input to a first input layer and a second input layer that are independent of each other, and feature quantities are calculated. Thus, the feature quantity calculation in one of the first and second input layers is not affected by the feature quantity calculation in the other input layer. In addition to feature extraction performed in the input layers, each of a first intermediate feature quantity calculation process and a second intermediate feature quantity calculation process is performed at least once in an intermediate layer that is shared by the first and second input layers. Thus, the feature quantities calculated from the first data and the second data in the respective input layers can be reflected in the intermediate feature quantity calculation in the intermediate layer. Consequently, pieces of data that belong to the same category and are acquired under different conditions can be appropriately learned.