Classification Network Training Using Multi-Layer Activation Maps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classification networks using Convolutional Neural Networks (CNNs) face challenges in maintaining accurate classification performance due to overfitting, particularly when trained with erroneous data, leading to deteriorated results.
Innovation Solution
An electronic apparatus is designed with a processor and memory to train a classification network by acquiring class scores and final loss values from multiple feature extraction layers, using activation maps and softmax loss values to control the network training, thereby improving classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If only class information is used for training the classification network, then the training process is simple and fast, but classification performance deteriorates due to overfitting
Solution Approach 1:
The patent segments the training process into multiple stages: first training with class information only for fast initial learning, then progressively adding position information and using activation maps from different feature extraction layers for refined training. This segmentation allows the model to achieve both fast training and high accuracy by combining simple initial training with detailed subsequent training.
Solution Approach 2:
The patent performs preliminary training using only class information to establish basic classification capabilities quickly. After this preliminary stage, the system progressively adds position information and uses activation maps from multiple feature extraction layers to refine the model, ensuring both fast initial progress and high final accuracy.
2Reliability
If position information is added to training data, then classification accuracy improves, but training complexity and computational cost increase
Solution Approach 1:
The patent performs preliminary training without position information to establish basic classification capabilities. Only after this preliminary stage does the system add position information and use activation maps from multiple feature extraction layers, thereby managing training complexity while improving accuracy progressively rather than all at once.
Solution Approach 2:
The patent introduces position information as an additional dimension to the training data, combining it with class information and activation maps from multiple feature extraction layers. This dimensional expansion allows the model to capture both what objects are and where they are, improving accuracy while managing complexity through structured integration.
3Productivity
If erroneous position data is used for training, then training data utilization is maximized, but classification performance deteriorates due to overfitting on incorrect positions
Solution Approach 1:
The patent uses activation maps from multiple feature extraction layers as feedback signals during training. These activation maps provide information about the network's internal representations and help distinguish between correct and erroneous position data, allowing the model to utilize training data effectively while avoiding overfitting to incorrect positions through the feedback from multiple layers.
Solution Approach 2:
The patent segments the training process into stages where different types of information are progressively introduced. By separating the training into initial class-only training followed by progressive addition of position information and multi-layer activation map utilization, the system can maximize data utilization while preventing overfitting through staged learning.
Data Source
AI summary
An electronic apparatus includes a memory for storing a classification network including a plurality of feature extraction layers. The electronic apparatus also includes a processor for acquiring a class score corresponding to an object, which is output from the classification network, by inputting a training image including the object to the classification network, acquiring a final loss value, based on a plurality of activation maps respectively output from the plurality of feature extraction layers and the class score, and controlling the classification network, based on the final loss value.