Reasoning as Remembering: The Theory and Practice of CBR

Kristian J. HammondAAAI 1992

Introduction

In the early 1990s, Artificial Intelligence researchers explored Memory-Based Reasoning (MBR) as a powerful paradigm where "reasoning" is conceptualized as "remembering" and matching similar past cases. However, the computational cost of finding the most similar case in databases containing millions of records was prohibitive for standard hardware. This paper, originally presented at AAAI-92, proposes a specialized hardware architecture—WSI-MBR—that leverages Wafer Scale Integration to achieve the massive parallelism required for real-time MBR on a desktop scale.

The Challenge of MBR

Memory-Based Reasoning is inherently parallel. To identify the "nearest neighbor" in a massive dataset, every record must ideally be compared to the target simultaneously. In 1992, the state-of-the-art for such tasks was the Connection Machine (CM-2), a room-sized supercomputer with 64,000 processing elements. The authors argued that for practical applications like protein structure prediction or national-scale data analysis, millions of parallel paths were necessary at a fraction of the cost and size of a supercomputer.

The WSI-MBR Architecture

The proposed solution uses Wafer Scale Integration (WSI), where an entire silicon wafer is used as a single system rather than being diced into individual chips. The architecture employs a hybrid analog/digital design to maximize density and speed.

The hybrid analog/digital memory cell architecture featuring analog multipliers for similarity scoring.

Digital circuits handle data storage and bit-wise comparisons, while analog circuits are used for "weight computing"—calculating similarity scores. By using a single transistor as an analog multiplier, the design achieves a much higher density than purely digital floating-point units. A global "threshold relaxing" mechanism allows the controller to identify the highest-scoring cells rapidly by lowering a threshold voltage until the best matches "fire."

Robustness and Fault Tolerance

A traditional barrier to WSI is manufacturing defects; a single flaw can often ruin a whole wafer. WSI-MBR overcomes this through "Defect Acceptance." Because MBR relies on the statistical aggregate of data, the system can simply ignore a small percentage of defective cells without significantly impacting overall accuracy. Furthermore, critical control paths utilize triple redundancy to prevent single-point hardware failures from disabling the entire wafer.

Letter accuracy vs. noise level demonstrates the system's robustness to weight variations.

Simulations using the NETtalk dataset confirmed this robustness. The system maintained over 80% letter accuracy even when subjected to ±100% noise in the similarity weights, proving that high-precision analog components are not a prerequisite for effective MBR.

Performance and Results

The WSI-MBR design achieved remarkable benchmarks for its time:

  • Massive Parallelism: 240,000 parallel records on a 5-inch wafer, with projections of 2 million records on an 8-inch wafer by 1995.
  • Extreme Speed: Inference times of approximately 66 nanoseconds, equivalent to 70 Tera FLOPS in digital terms.
  • Fault Tolerance: At typical defect densities of the era, only 0.02% of cells would be affected, leading to negligible performance loss.

Conclusion

By aligning hardware architecture with the algorithmic strengths of Memory-Based Reasoning, the authors demonstrated a path toward "laptop-sized supercomputers." While bottlenecks in data fetching and the lack of flexibility in hardwired logic remain challenges, WSI-MBR represents a significant milestone in specialized AI hardware, proving that robustness to noise and defects can be a primary design feature rather than a limitation.

Summary updated

BibTeX
@inproceedings{hammond1992reasoning,
  author = {Kristian J. Hammond},
  title = {{Reasoning as Remembering: The Theory and Practice of CBR}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure