An Optimally Efficient Limited Inference System

Lokendra Shastri, Venkat AjjanagaddeAAAI 1990

TL;DR

This paper introduces a novel connectionist system that performs a restricted but powerful class of inferences with optimal efficiency, independent of the knowledge base size. It proposes that achieving human-like real-time reasoning requires limiting inference capabilities rather than seeking universal, intractable solutions, offering a path towards efficient AI systems.

What Problem Does This Paper Address?

The paper tackles the "inference paradox" in Artificial Intelligence: human cognition, particularly language understanding, appears to perform complex reasoning tasks with remarkable speed and efficiency, yet generalized logical inference is computationally intractable. This suggests that humans don't employ general-purpose reasoning but rather a "limited" yet "rich" class of inferences performed reflexively. The challenge is to understand and replicate this efficient, limited inference capability. The authors propose that by carefully characterizing and restricting the types of inferences and the expressiveness of the knowledge representation, it's possible to build AI systems that are both powerful and highly efficient, achieving performance that scales independently of the knowledge base size.

Methods

The core of the proposed system is a massively parallel connectionist network. Key methodological innovations include:

  • Oscillatory Nodes: The system utilizes simple "phase-sensitive binary threshold units" (p-btu) and "temporal and nodes" (r-and nodes) that produce oscillatory outputs. The timing (phase) of these oscillations is crucial for representing and propagating information.
  • Dynamic Variable Binding: A significant technical challenge addressed is the run-time creation and propagation of variable bindings in a parallel network. The system achieves this through the rhythmic patterns of activation and the precise timing of node firings, effectively creating dynamic links between variables and their instantiations.
  • Restricted Inference Class: The system is designed to handle a specific class of first-order sentences and queries. This restriction is key to achieving optimal efficiency. Constraints include limitations on the number of distinct constants in a query and how variables can be used within rules.
  • Encoding Scheme: Knowledge is encoded by representing predicates and their arguments as nodes in the network. Rules are represented by links connecting these nodes, forming a directed inferential dependency graph. Facts are encoded as small networks that match specific patterns of activation.

Datasets

Not discussed in the paper. The paper focuses on the theoretical framework and realization of the system rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Efficiency Through Restriction: The most profound insight is that achieving optimal efficiency in reasoning might not come from more powerful general algorithms, but from carefully limiting the expressiveness and inferential capabilities of the system. This is a counter-intuitive but powerful idea for AI design.
  • The Power of Timing: The system leverages the temporal dynamics (phase and rhythm) of oscillatory nodes to represent and manipulate variable bindings, a departure from purely static symbolic manipulation. This connectionist approach offers a biologically plausible mechanism for fast reasoning.
  • Symbiotic Relationship: The paper emphasizes the interconnectedness of expressiveness, inferential power, representation choice, and computational models. Optimizing one often requires trade-offs with others, and this system finds a sweet spot for efficient inference.
  • Reflexive Reasoning as a Model: The concept of "reflexive reasoning" – fast, effortless, and seemingly automatic cognitive processes – is proposed as a more accurate model for human-like AI than general-purpose logical deduction.

Key Results

  • Optimal Efficiency: The system answers yes-no and wh-queries in time proportional to the length of the shortest derivation of the query, and is independent of the size of the knowledge base (|KB|).
  • Linear Space Complexity: The knowledge representation requires space that is only linear in the size of the knowledge base (|KB|).
  • Soundness (with conditions): The system is sound provided certain conditions are met, related to the number of constants in queries and specific constraints on rule usage.

Limitations and Future Work

  • Limited Inference Class: The system's primary limitation is its restricted class of inference. It cannot handle all first-order logic queries and has specific constraints on rule structure and query formulation.
  • Incompleteness: The paper acknowledges that the system is incomplete, meaning it may not answer all valid queries. The conditions provided characterize this incompleteness.
  • Future Work: The authors suggest extensions to interface with specialized modules like semantic networks or IS-A hierarchies to enhance expressiveness. They also mention allowing a limited use of function terms in rules. Further research could explore the biological plausibility of the underlying mechanisms and extend the system's capabilities to handle more complex logical constructs.

Summary updated

BibTeX
@inproceedings{shastri1990optimally,
  author = {Lokendra Shastri and Venkat Ajjanagadde},
  title = {{An Optimally Efficient Limited Inference System}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}

Loading related content...

Figure