Group-based local adaptive deep multiple kernel learning with lp norm.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Deep multiple kernel learning (DMKL) offers superior classification performance over shallow methods.
• Existing DMKL methods struggle with global parameter optimization and fail to address inter-class correlation and intra-class diversity.
• Accurate classification on diverse datasets remains a challenge for current DMKL approaches.

Purpose of the Study:

• To introduce a novel group-based local adaptive deep multiple kernel learning (GLDMKL) method.
• To enhance classification accuracy by addressing limitations of existing DMKL techniques.
• To improve adaptability and handle complex datasets more effectively.

Main Methods:

• The proposed GLDMKL method utilizes multiple kernel k-means clustering to group samples.
• Adaptive deep multiple kernel learning is applied within each local group, allowing for flexible network structures.
• Support Vector Machine (SVM) parameters (α) and local kernel weights (β) are optimized iteratively, with kernel weights constrained by the lp norm.
• Hyperparameters are tuned using grid search.

Main Results:

• Experiments on UCI and Caltech 256 datasets show GLDMKL outperforms other DMKL methods in classification accuracy.
• The method demonstrates particular effectiveness on datasets with complex data structures.
• GLDMKL successfully adapts to local data characteristics, improving model generalization.

Conclusions:

• The GLDMKL method provides a robust framework for improving classification accuracy in deep multiple kernel learning.
• Local adaptation and group-based learning effectively address the limitations of global parameter optimization in DMKL.
• GLDMKL offers a promising advancement for handling complex and diverse datasets in machine learning applications.