Introduction
In the early 1980s, the field of Artificial Intelligence faced a significant architectural dilemma regarding expert systems. "First-generation" systems were often criticized for being "shallow," relying on flat databases of situation-action rules (e.g., "If symptom X, then disease Y"). While computationally efficient, these systems lacked "common sense" and the ability to explain the underlying rationale for their rules. Conversely, "deep" systems attempted to model the world from first principles using causality and anatomy. While robust, reasoning from scratch for every case is computationally expensive and does not reflect the specialized clinical patterns used by human experts.
In their 1982 paper, Chandrasekaran and Mittal propose a middle path: compiled knowledge. They argue that for tasks like medical diagnosis, we can build efficient systems that "pre-digest" deep causal reasoning into a specialized, hierarchical structure.
The MDX Architecture
The authors introduce MDX, a medical diagnosis system designed to capture the power of deep knowledge without the overhead of real-time first-principles reasoning. Its core approach rests on two pillars: the Specialist Hierarchy and the Establish-Refine Strategy.
The Specialist Hierarchy
Knowledge in MDX is organized into a tree of "specialists." At the top level, a general node like "Internist" oversees the process, branching down into more specific domains such as "Liver" or "Heart." These, in turn, branch into specific conditions like "Hepatitis" or "Cholestasis."
Establish-Refine Strategy
Problem-solving moves top-down through this hierarchy. A node first attempts to "establish" itself—determining if the current case falls within its domain. If successful, it "refines" the hypothesis by passing control to its children. If a node fails to establish itself, it prunes the entire branch, drastically increasing efficiency by ignoring irrelevant data.
The "deep" aspect of this system is integrated during the design phase. The rules within these specialists are "compiled" from underlying (U) knowledge about anatomy and physiology. For instance, the deep physiological knowledge that "bile ducts are elastic tubes that dilate when blocked" is compiled into a specific diagnostic rule: "If X-ray shows dilation, conclude Extrahepatic Obstruction (EHO)."
Key Insights
- Task-Specific Knowledge: The authors argue that while "deep" knowledge is often general, "diagnostic" knowledge is a specialized transformation of that information. One "compiles" general knowledge into a form optimized for the specific task of diagnosis.
- Efficiency through Pruning: The hierarchical structure ensures the system avoids looking at irrelevant data. If the "Heart" specialist rules itself out, the system never considers specific cardiac diseases.
- Dual-Layer Explanation: The paper distinguishes between two types of "Why?":
- Type 1 (Patient-specific): "Why do you think this patient has a stone?" This is answered by the compiled rules triggered during the inference process.
- Type 2 (Domain-specific): "Why does a dilated duct indicate an obstruction?" This is answered by pre-stored text representing the "deep" rationale used during the initial compilation phase.
Results and Limitations
The MDX system demonstrated that it could successfully navigate complex diagnostic paths, such as identifying Extrahepatic Obstruction, using only its compiled hierarchy. This proof-of-concept showed that one can achieve the performance of a "deep" system with the speed of a "surface" system by carefully designing the knowledge hierarchy.
However, the approach has limitations:
- Compilation Burden: Currently, "compiling" the diagnostic structure from underlying knowledge is a manual task for human experts.
- Rigidity: Knowledge compiled for diagnosis cannot be easily repurposed for teaching or therapy without re-compilation.
- The Fallback Problem: For extremely rare or "combinatorial" cases, a compiled system may fail. In such instances, a system must revert to "run-time compilation"—reasoning from first principles because no pre-built shortcut exists.
The authors conclude that "deep" knowledge is most valuable during the design and explanation phases of an AI, rather than being strictly necessary for every step of the inference phase.
Summary updated