Deep Dictionary Learning: A PARametric NETwork Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep dictionary learning utilizes multiple dictionaries for multi-scale image analysis.
• Existing methods may not optimally integrate feature extraction with classification objectives.

Purpose of the Study:

• To propose a hierarchical synthesis dictionary learning method for image classification.
• To enhance feature representation by integrating reconstruction and classification objectives.
• To improve model adaptability and adversarial robustness.

Main Methods:

• Learning a hierarchy of synthesis dictionaries trained with a classification objective.
• Extracting sparse features by minimizing reconstruction loss in each layer.
• Regularizing classification using reconstruction objectives to incorporate source signal information.

Main Results:

• Performance improves with increased model depth (more layers).
• The hierarchical model demonstrates significantly lower fooling rates against adversarial perturbations.
• The method achieves competitive classification performance on benchmark datasets compared to similar-sized CNNs.

Conclusions:

• Hierarchical dictionary learning offers a scalable and adaptable approach for image classification.
• Integrating reconstruction loss enhances feature interpretability and adversarial resilience.
• The proposed method provides a valuable alternative to standard deep learning architectures for image classification tasks.