Neural Network Model Training for Complex Characteristic Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural network models struggle to properly reflect and predict complex characteristics, as they are limited by the requirement that the total sum of result values must be 100%, making it difficult to accurately classify and localize complex characteristics such as gender and ethnicity.
Innovation Solution
A neural network model training method that includes a convolution layer, a pooling layer, and class-specific fully connected layers, where class-specific observation maps are calculated and an observation loss is computed based on these maps, allowing for the back-propagation of loss to improve classification and localization of complex characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a conventional neural network model is used to classify characteristics, then the total sum of result values is constrained to be 100%, but this prevents proper reflection of complex characteristics where multiple results can occur simultaneously
Solution Approach 1:
The patent segments the classification task by introducing separate classification heads for different characteristic types (e.g., gender classifier, ethnicity classifier) instead of using a single constrained classification layer. Each head independently processes its specific characteristics, allowing multiple characteristics to be predicted simultaneously without the 100% sum constraint limiting any single characteristic's prediction accuracy.
Solution Approach 2:
The patent transitions from a single-dimensional probability distribution (where sum must be 100%) to a multi-dimensional prediction space where each characteristic can be predicted independently. By adding separate classification branches for different characteristic types, the model can output multiple results simultaneously without normalization constraints interfering with any single characteristic's prediction.
2Measurement precision
If class-specific fully connected layers are added to handle multiple classes, then classification accuracy for complex characteristics improves, but model complexity increases
Solution Approach 1:
The patent implements a shared convolutional backbone that serves multiple classification heads simultaneously. The convolutional layers extract shared features that are then distributed to different classification heads (gender, ethnicity, etc.), allowing the same feature extraction infrastructure to serve multiple classification purposes without requiring separate convolutional networks for each characteristic type.
Solution Approach 2:
The patent segments the classification function into separate heads while sharing the feature extraction backbone. This segmentation allows each classification head to be optimized for its specific characteristics without requiring the entire model to be redesigned, thus managing complexity by isolating classification-specific parameters while reusing shared convolutional weights.
Data Source
AI summary
A neural network model training method and an apparatus for complex characteristic classification and common localization are proposed. In the method, a neural network model includes: a convolution layer for performing a convolution operation on an input image by using a convolution filter; a pooling layer for performing pooling on an output of the convolution layer; and class-specific fully connected layers respectively corresponding to classes into which complex characteristics are classified and outputting values obtained by multiplying an output of the pooling layer by class-specific weights (Wfc(Tt). The method includes: (a) inputting the input image to the convolution layer; (b) calculating class-specific observation maps for respective classes on the basis of the output of the convolution layer; (c) calculating an observation loss (Lobs) common to the classes on the basis of the class-specific observation maps; and (d) back-propagating a loss based on the observation loss to the neural network model.