TL;DR
This paper introduces LREX, a novel algorithm for extracting symbolic rules from Radial Basis Function (RBF) networks. By combining two distinct rule extraction methods (mREX and hREX), LREX provides a more comprehensive understanding of the RBF network's internal workings, addressing limitations of previous approaches.
What Problem Does This Paper Address?
Neural networks, particularly RBF networks, are powerful tools for complex pattern recognition and classification. However, their "black box" nature makes it difficult to understand how they arrive at their decisions. This lack of interpretability hinders their adoption in critical applications and limits knowledge discovery. Extracting symbolic rules from these networks can bridge this gap, making their decision-making process transparent and understandable. Previous attempts at rule extraction from RBF networks faced challenges due to nonlinear functions and high input dimensionality, leading to issues like "shared" hidden units that complicate rule generation.
Methods
The paper proposes the LREX (Local Rule EXtraction) algorithm, which tackles the challenges of rule extraction from RBF networks through a two-pronged approach:
- mREX (Input-to-output mapping): This module extracts IF-THEN rules by directly mapping input patterns to output classes. It leverages the RBF network's structure, specifically the hidden unit assignments to output classes, to generate rules. This method is effective for simpler data sets but can struggle with hidden unit sharing in more complex scenarios.
- hREX (Hidden unit analysis): This module takes a different approach by analyzing the internal structure and relationships formed by the hidden units. It uses quantization and clustering to abstract the network's operation, providing insights into how hidden units contribute to classification. hREX is particularly useful for understanding hidden unit sharing and can generate rules that are more comprehensible due to user-defined limits on rule complexity.
LREX combines the strengths of both mREX and hREX to offer a more complete picture of the RBF network's knowledge.
Datasets
The paper evaluates LREX on a variety of datasets, including:
- Standard benchmark datasets: Xor (binary and continuous), Iris, Vowell (Peterson and Deterding), Protein (yeast and ecoli), Credit (Japanese and Australian), Diabetes (Pima), Monks 1, Sonar.
- Newly introduced vibration datasets: Vibration 1 and Vibration 2, derived from projects monitoring the health of industrial machinery.
These datasets represent a range of complexities, including varying numbers of examples, input features, and classes, and some contain continuous or discrete data, with some also having missing data.
Key Insights and Ideas
- RBF networks are amenable to rule extraction: The paper's central claim is that the architecture of RBF networks, with their Gaussian functions, is inherently suitable for extracting symbolic rules.
- Two-level extraction is better than one: Combining mREX (input-output mapping) and hREX (internal structure analysis) provides a more comprehensive and insightful rule set than either method alone.
- Hidden unit sharing is a key challenge: The paper highlights that hidden unit sharing across classes in complex RBF networks complicates direct rule extraction, necessitating the hREX approach.
- Trade-off between rule size and comprehensibility: hREX allows users to control the number of rules, offering a trade-off between rule set size (and thus comprehensibility) and accuracy.
- LREX offers valuable insights: The extracted rules from LREX can reveal information about the original RBF network's internal structure, identify potential issues with training, and provide a transparent, rule-based classifier.
Key Results
The paper compares LREX (mREX and hREX) against a competing system, RULEX, and the original RBF network performance.
- Accuracy: LREX generally achieves comparable or better accuracy than the original RBF networks and RULEX on many datasets. For instance, on the Xor (binary) dataset, both mREX and hREX achieve 100% accuracy, matching the original RBF network and RULEX.
- Rule Set Size: LREX generates a larger number of rules compared to RULEX, which produces more compact rule sets. However, LREX's rules are derived directly from the network's structure, offering a different kind of interpretability.
- Robustness: LREX demonstrates its ability to extract rules even when RBF networks have poor performance on certain classes, providing insights into hidden unit sharing.
Limitations and Future Work
- RULEX's compactness: The paper acknowledges that RULEX extracts more compact rule sets, which can be advantageous for comprehensibility.
- LREX rule count: The number of rules generated by mREX is dependent on the number of hidden units, which can sometimes lead to a large number of rules that might obscure comprehensibility.
- Future work: The authors suggest exploring the application of LREX to other types of neural networks and further investigating the relationship between rule complexity and accuracy. They also mention the potential for combining LREX with genetic algorithms for further optimization.
Summary updated