Use of a Deep Belief Network for Small High-Level Abstraction Data Sets Using Artificial Intelligence with Rule...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Deep neural networks (DNNs) trained by deep belief networks (DBNs) offer high learning capabilities but lack interpretability.
• Rule extraction algorithms, like recursive-rule eXtraction (Re-RX) with J48graft, provide interpretability but may struggle with complex DNNs.
• A gap exists between the performance of DBN-trained DNNs and the interpretability of rule extraction methods.

Purpose of the Study:

• To propose a novel method for transferring weights from DBN-trained DNNs to backpropagation neural networks (BPNNs).
• To extract accurate and interpretable classification rules from rating category datasets using the proposed method.
• To address the challenge of bridging the interpretability gap in deep learning models.

Main Methods:

• A simple weight transfer method from deep belief network (DBN) trained deep neural networks (DNNs) to backpropagation neural networks (BPNNs).
• Integration of this transfer method into the recursive-rule eXtraction (Re-RX) algorithm with J48graft.
• Application and validation of the proposed rule extraction method on the Wisconsin Breast Cancer Data Set (WBCD), Mammographic Mass Data Set, and Dermatology Dataset.

Main Results:

• The proposed method successfully extracted accurate and concise classification rules from DBN-trained DNNs on three distinct datasets.
• Demonstrated the effectiveness of the weight transfer mechanism for rule extraction.
• Achieved high accuracy and interpretability in the extracted classification rules.

Conclusions:

• The developed method effectively bridges the gap between the high learning capacity of DBNs and the interpretability of rule extraction algorithms.
• The approach facilitates the extraction of understandable classification rules from complex deep learning models.
• This technique offers a promising solution for enhancing the interpretability of deep neural networks in various applications.