CNN filter sizes, effects, limitations, and challenges: An exploratory study.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) commonly utilize small filters (e.g., 3x3).
• A prevailing belief suggests small filters yield superior results, potentially overlooking the benefits of larger filters.
• Limited research exists on optimizing CNN performance with large filters.

Purpose of the Study:

• To challenge the conventional wisdom favoring small filters in CNNs.
• To explore and highlight the untapped potential of large filters in CNN architectures.
• To encourage broader investigation into the impact of filter sizes on model performance.

Main Methods:

• Analysis of existing research on CNN architectures and filter sizes.
• Exploration of limitations and challenges associated with using large filters.
• Comparative study of small versus large filter impacts across different CNN models.

Main Results:

• Identified a significant bias towards small filters (3x3) in current CNN research.
• Uncovered four distinct opportunities for performance enhancement using large filters.
• Demonstrated the potential for large filters to improve CNN model capabilities.

Conclusions:

• The efficacy of large filters in CNNs is an under-explored area.
• Advancements in computational power and image sizes reduce previous barriers to large filter adoption.
• Further research into large filters is crucial for advancing CNN performance and applications.