Hierarchical CNN Training with Skip Connections and Progressive Data Sets

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing learning models for hierarchical image classification face efficiency decreases as the number of classes increases, leading to a decline in accuracy during complex training tasks.

Innovation Solution

A learning model configuration that includes a convolutional neural network with a main network and sub network, utilizing skip connections where higher-level classifiers connect to lower-level classifiers, and generates training data sets through random addition and deletion of elements to enhance accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the number of classes in each layer increases and the classification becomes more complex, then the hierarchical classification capability is improved, but the processing efficiency for training the learning model decreases

Engineering Contradiction:
Improvehierarchical classification capabilityVSAvoidtraining processing efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent divides the classification task into multiple hierarchical layers, where each layer handles a specific level of classification (e.g., category level, subcategory level, product level). This segmentation allows the model to process complex hierarchical classification by breaking it down into manageable steps, improving both classification capability and training efficiency through layered processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs dynamic training strategies where the learning model adapts its processing based on the complexity of the classification task. The model dynamically adjusts its behavior during training to handle increasing numbers of classes and complexity levels, maintaining processing efficiency while improving hierarchical classification capability through adaptive learning mechanisms.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If the number of classes in each layer increases and the classification becomes more complex, then the hierarchical classification capability is improved, but the accuracy of the learning model decreases

Engineering Contradiction:
Improvehierarchical classification capabilityVSAvoidclassification accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

By segmenting the classification into hierarchical layers, each layer focuses on specific classification objectives with optimized parameters. This segmentation prevents accuracy degradation by ensuring that each layer achieves high accuracy for its specific level before proceeding to more complex levels, thereby maintaining overall accuracy despite increasing classification complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent incorporates feedback mechanisms that monitor and adjust the training process based on performance at each hierarchical level. This feedback allows the model to identify and correct accuracy issues at specific levels, maintaining high classification accuracy even as the number of classes and overall complexity increase across multiple layers.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4216114B1Information processing device, information processing method, and program
Publication Date: 2025.08.13 RAKUTEN GROUP INC
  • EP4216114B1 patent drawingFigure 1
  • EP4216114B1 patent drawingFigure 2
  • EP4216114B1 patent drawingFigure 3

AI summary

An information processing apparatus includes: a generating means for generating, from a data group including training data elements given different correct labels, multiple training data sets in such a manner that the number of included training data elements changes successively, by successively selecting one or more training data elements from the data group; and a training means for repeatedly training a learning model for machine learning by applying the training data sets to the learning model.