A Hybrid Connectionist, Symbolic Learning System

Lawrence O. Hall, Steve G. RomaniukAAAI 1990

Introduction

The divide between symbolic AI (rule-based expert systems) and connectionist AI (neural networks) has long defined the landscape of artificial intelligence. While symbolic systems offer transparency and logic, they suffer from the "knowledge acquisition bottleneck." Conversely, neural networks excel at learning from data but often function as "black boxes" that struggle with symbolic structures. This paper introduces SC-net, a hybrid system designed to bridge this gap by learning rules from data in a single pass and representing them through fuzzy logic.

The SC-net Architecture

At its core, SC-net utilizes fuzzy logic cells rather than standard sigmoidal neurons. These cells perform operations such as fAND (minimum), fOR (maximum), and fNOT (negation), with the cell's bias determining the specific logic applied. A distinctive feature of the architecture is its treatment of uncertainty. By defining an activation of 0.5 as "unknown" (UK), the system can explicitly differentiate between false evidence and a lack of knowledge.

Figure 1: The SC-net architecture illustrating positive/negative collectors and the 'unknown' (UK) node.

As shown in Figure 1, the network uses "collector" nodes to aggregate evidence. This structure allows the system to maintain the interpretability of a rule-based system while leveraging the parallel processing of a connectionist model.

Learning via Cell Recruitment

Unlike traditional backpropagation, which requires thousands of epochs to adjust weights, SC-net employs a "Recruitment of Cells" algorithm. This approach allows for incremental, one-pass learning:

  1. Evaluation: The system determines if the current network can explain a new example.
  2. Fine-tuning: If the error is minimal, existing biases are adjusted.
  3. Recruitment: If the example is "unknown" to the network, it recruits a new hidden cell from a "conceptual pool" to represent the new knowledge.

Handling Symbolic Variables

SC-net effectively manages attribute-value pairs, such as Age = 15. It represents these variables using groups of connected cells that define fuzzy intervals. This allows the network to process continuous data within a symbolic framework—for instance, recognizing that a 15-year-old is a "teenager" with high confidence while also potentially belonging to overlapping categories.

Extracted figure from page 5 of A Hybrid Connectionist, Symbolic Learning System

Experimental Results

The authors validated SC-net across several domains:

  • Iris Dataset: The system achieved a 96% accuracy rate (0.04 error), performing competitively with established pattern recognition techniques.
  • Medical Knowledge Bases: Using the FEVERS and GEMS databases from EMYCIN, SC-net demonstrated high fidelity to human-defined rules, with certainty factors closely matching the original expert systems.
  • Industrial Application: The system successfully identified defects in CMOS semiconductor manufacturing, proving its utility in real-world diagnostic tasks.

Conclusion

SC-net represents a significant step toward interpretable machine learning. By combining the efficiency of one-pass learning with the transparency of fuzzy logic, it addresses the limitations of both symbolic and connectionist paradigms. Future work may explore how this recruitment strategy scales to larger datasets and potential mechanisms for pruning redundant cells to optimize network topology.

Summary updated

BibTeX
@inproceedings{hall1990hybrid,
  author = {Lawrence O. Hall and Steve G. Romaniuk},
  title = {{A Hybrid Connectionist, Symbolic Learning System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure