Learning Hierarchical Representations with Spike-and-Slab Inception Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) with inception modules show strong performance.
• Basic CNNs capture only linear, univariate features, limiting expression and feature mining.
• Deepening networks leads to parameter redundancy and overfitting.

Purpose of the Study:

• To improve CNNs for enhanced image recognition capabilities.
• To address the limitations of univariate feature extraction in CNNs.
• To explore novel architectures that improve feature expression without increasing parameters.

Main Methods:

• Introduced spike-and-slab units into a modified inception module.
• Enabled the model to capture dual latent variables, including average and covariance information.
• Evaluated the enhanced model on several image recognition tasks.

Main Results:

• Dual variable operations were successfully integrated into inception modules.
• The enhanced model demonstrated improved robustness to image intensity variations.
• Excellent results were achieved across various image recognition tasks.

Conclusions:

• The proposed method enhances CNN performance by capturing richer feature representations.
• Integrating spike-and-slab units offers a robust solution for image recognition challenges.
• This approach provides a promising direction for future deep learning architectures.