Visual Representation and Classification by Learning Group Sparse Deep Stacking Network.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep stacking networks (DSNs) utilize simplified neural network modules (SNNMs) for classification.
• The independence assumption in SNNMs limits learning local dependencies crucial for accurate classification.
• Real-world classification often involves class-specific data representations that can be grouped.

Purpose of the Study:

• To address the limitations of SNNMs in capturing local dependencies.
• To enhance DSNs by incorporating group sparsity for improved classification performance.
• To develop a novel group sparse DSN (GS-DSN) model.

Main Methods:

• Proposed two group sparse SNNM modules by mixing L1 and L2 norms.
• The first module captures local dependencies by grouping hidden units.
• The second module clusters samples by grouping representations of different classes.

Main Results:

• A group sparse DSN (GS-DSN) was constructed by stacking the proposed modules.
• GS-DSN demonstrated superior performance compared to existing classification methods.
• Achieved state-of-the-art accuracy of 99.1% on the 15-Scene dataset.

Conclusions:

• The proposed group sparse SNNM modules effectively learn local dependencies and class-specific representations.
• GS-DSN significantly outperforms traditional DSNs and other relevant classification methods.
• The findings highlight the efficacy of group sparsity in enhancing deep learning models for classification tasks.