An Empirical Evaluation of Rule Extraction from Recurrent Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Linguistics

Background:

• Rule extraction from black box models is crucial for validation in sensitive fields like credit scoring and medical diagnosis.
• Highly nonlinear and recursive models, such as recurrent neural networks (RNNs), present significant challenges for rule extraction.
• Second-order RNNs were trained to recognize Tomita grammars, a benchmark for formal language recognition.

Purpose of the Study:

• To investigate the feasibility and stability of extracting production rules from trained second-order RNNs.
• To evaluate the performance of extracted rules compared to the original RNN models.

Main Methods:

• Training second-order recurrent neural networks (RNNs) on the Tomita grammars.
• Applying rule extraction techniques to the trained RNNs to derive production rules.
• Comparing the performance of the extracted rules against the trained RNNs on the grammar recognition task.

Main Results:

• Production rules were successfully and stably extracted from the trained second-order RNNs.
• In specific instances, the extracted rules demonstrated superior performance compared to the original RNNs in recognizing the Tomita grammars.

Conclusions:

• Rule extraction from complex RNNs is achievable and can yield interpretable models.
• Extracted rules can potentially offer improved performance over the black box models they originate from, particularly in grammar recognition tasks.