Replacing pooling functions in Convolutional Neural Networks by linear combinations of increasing functions.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional Convolutional Neural Networks (CNNs) utilize max or mean pooling for feature reduction.
• These pooling methods have limitations, ignoring data dependencies and lacking a universally superior choice.
• Problem-dependent selection of pooling functions hinders optimal performance in CNN architectures.

Purpose of the Study:

• To develop and evaluate novel pooling strategies that enhance feature extraction in CNNs.
• To address the limitations of traditional pooling by incorporating combined and generalized functions.
• To introduce the CombPool layer as an alternative to standard pooling mechanisms.

Main Methods:

• Replaced traditional pooling layers with alternative functions, including linear combinations of order statistics.
• Extended and generalized the Sugeno integral for broader applicability in pooling strategies.
• Implemented and tested the proposed CombPool layer across three CNN architectures of varying complexity.

Main Results:

• Empirical evidence demonstrates that linear combinations of pooling functions outperform traditional methods on multiple datasets.
• Combinations involving the Sugeno integral or its generalizations consistently yielded the best performance.
• The CombPool layer proved effective in improving feature extraction across different CNN architectures.

Conclusions:

• The proposed CombPool strategy, utilizing combinations of pooling functions, offers significant advantages over traditional methods.
• Generalized Sugeno integrals show strong potential for enhancing feature extraction in deep learning models.
• CombPool represents a robust and adaptable pooling solution applicable to a wide range of CNN architectures.