Enhancing the accuracies by performing pooling decisions adjacent to the output layer

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models commonly use max-pooling (MP) operators throughout feedforward architectures for classification tasks.
• The placement and type of pooling operators significantly influence model performance and feature selection.

Purpose of the Study:

• To investigate the impact of pooling operator placement, specifically adjacent to the last convolutional layer, on classification accuracy.
• To compare the performance of advanced VGG (A-VGG) architectures with modified pooling strategies against standard deep learning models.

Main Methods:

• Utilized the CIFAR-10 database for evaluating deep learning model performance.
• Implemented and tested various advanced-VGG (A-VGGm) architectures with modified pooling strategies near the output layer.
• Compared results with established architectures like VGG16 and Wide-ResNet16.

Main Results:

• Pooling decisions adjacent to the last convolutional layer significantly enhanced classification accuracies.
• A-VGG8 achieved higher accuracy than VGG16, and A-VGG13/A-VGG16 matched Wide-ResNet16 performance.
• Modifications to fully connected (FC) layers also yielded comparable accuracies, indicating the primary impact of output-adjacent pooling.

Conclusions:

• The strategic placement of pooling operators, particularly average pooling, adjacent to the output layer is crucial for improving deep learning classification.
• This approach effectively trains more influential input-output routes, leading to superior performance compared to traditional deep architectures.
• Results suggest a reevaluation of existing deep learning architectures and their reported accuracies using this proposed pooling strategy.