Multi-Channel Model Training for Robust Irregular Image Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional learning models lack robustness when irregular input images are encountered due to overfitting, leading to deteriorated inference accuracy.
Innovation Solution
An information processing apparatus and method that involves setting multiple input channels for different features, randomly selecting and deactivating some channels during training, and using pre-training with additional feature channels to enhance robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If dropout technique is used to prevent overfitting, then overlearning is prevented, but the learning model lacks robustness when irregular input images are encountered
Solution Approach 1:
The input channels are segmented into multiple groups (first input channel group for color features, second input channel group for text features). By randomly selecting and deactivating entire groups rather than individual channels, the model learns robust feature representations that generalize better to irregular inputs while maintaining inference accuracy.
Solution Approach 2:
The dropout mechanism is made dynamic by randomly selecting different groups of input channels for each training iteration. This dynamic selection prevents the model from over-relying on any single feature group, thereby improving robustness without sacrificing inference accuracy on regular inputs.
2Adaptability or versatility
If conventional dropout is applied to convolutional layers, then overfitting is prevented, but the model cannot handle irregular input data effectively
Solution Approach 1:
Input channels are divided into semantically meaningful groups (color features, text features). This segmentation allows the dropout mechanism to work at the group level, enabling the model to adapt to irregular inputs by learning from alternative feature groups while maintaining overall inference accuracy.
Solution Approach 2:
Each input channel group serves multiple functions: it provides specific feature information (color or text) and acts as a backup when other groups are deactivated. This multi-functionality enables the model to handle irregular inputs effectively while preserving inference accuracy across different input types.
Data Source
AI summary
Disclosed herein is an information processing apparatus including: a training data acquisition unit configured to acquire training data; an input channel setting unit configured to set a plurality of input channels for inputting different features, respectively, from the training data acquired by the training data acquisition unit to a learning model to be trained with the training data; and a training unit configured to, in each training, randomly select, among the plurality of input channels set by the input channel setting unit, at least a part of the input channels, and deactivate an input from the selected input channel to the learning model to repeat training of the learning model.