Improving CNN Performance Accuracies With Min-Max Objective.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are widely used for image classification and face verification.
• Improving CNN performance often requires increased network complexity or computational cost.
• Existing methods may not efficiently handle large-scale training data.

Purpose of the Study:

• To propose a novel Min-Max objective to enhance CNN performance accuracies.
• To achieve improved accuracy without increasing network complexity.
• To enable efficient training on large-scale datasets.

Main Methods:

• Applying a Min-Max objective to a layer below the output layer during CNN training.
• Ensuring minimum within-manifold distance and maximum between-manifold distances for feature maps.
• Utilizing an incremental minibatch training procedure for large-scale data.

Main Results:

• Significant improvements in performance accuracies for CNN models.
• Demonstrated effectiveness on benchmark datasets for image classification and face verification.
• Comparable performance with models trained without the Min-Max objective.

Conclusions:

• The Min-Max objective is a general and computationally efficient method for improving CNNs.
• The proposed training procedure effectively handles large-scale data.
• This approach offers a valuable technique for enhancing deep learning model performance.