Mesarovician Abstract Learning Systems

Tyler Cody

Abstract

The solution methods used to realize artificial general intelligence (AGI) may not contain the formalism needed to adequately model and characterize AGI. In particular, current approaches to learning hold notions of problem domain and problem task as fundamental precepts, but it is hardly apparent that an AGI encountered in the wild will be discernable into a set of domain-task pairings. Nor is it apparent that the outcomes of AGI in a system can be well expressed in terms of domain and task, or as consequences thereof. Thus, there is both a practical and theoretical use for meta-theories of learning which do not express themselves explicitly in terms of solution methods. General systems theory offers such a meta-theory. Herein, Mesarovician abstract systems theory is used as a super-structure for learning. Abstract learning systems are formulated. Subsequent elaboration stratifies the assumptions of learning systems into a hierarchy and considers the hierarchy such stratification projects onto learning theory. The presented Mesarovician abstract learning systems theory calls back to the founding motivations of artificial intelligence research by focusing on the thinking participants directly, in this case, learning systems, in contrast to the contemporary focus on the problems thinking participants solve.

Introduction

Current AI research is dominated by a "problem-solving" paradigm. We define a domain (input distribution) and a task (mapping inputs to outputs) and then build a specific solution method to minimize an error function. However, this approach is increasingly viewed as insufficient for the development and characterization of Artificial General Intelligence (AGI) for several reasons. First, AGI "in the wild" may not be easily decomposable into neat domain-task pairs due to its unbounded complexity. Second, intelligence may be an emergent property of a "system of systems" that cannot be reduced to a single solution method. Finally, AGI will likely influence and be influenced by its environment so deeply that the boundary between the "intelligent agent" and its "context" becomes blurred.

To address these challenges, this paper proposes a meta-theory—a way to model learning systems without dependency on specific algorithmic implementations. By leveraging Mesarovician Abstract Systems Theory (AST), we can view learning as a formal relation on sets, independent of the specific math of a neural network or reinforcement learning agent.

Methodology: Abstract Learning Systems Theory (ALST)

The core of the proposed methodology is the extension of AST into Abstract Learning Systems Theory (ALST). This involves stratification—organizing the assumptions and components of a learning system into three hierarchical levels of increasing complexity:

  1. Elementary Level: Treats learning as a simple input-output relation ($S \subset \times \{D, \mathcal{X}, \mathcal{Y}\}$), focusing on basic systemic properties like stability.
  2. Cascade Level: Decomposes the system into an inductive part (Algorithm $A$ and Parameters $\Theta$) and a functional part (Hypothesis $H$). This allows for the study of how the system interprets data to form internal representations.
  3. Goal-Seeking Level: Unpacks the inductive part to reveal internal feedback loops, consistency relations ($G, E$), and values ($V$). This level acknowledges that learning is an active, seeking process.
The three levels of stratification in Abstract Learning Systems Theory (ALST) and their corresponding AST diagrams.

Key Insights and Theoretical Mapping

By treating learning as a formal system ($S$), we can apply engineering concepts like "controllability" and "stability" to AGI in a way that traditional machine learning theory often misses. A significant contribution of this framework is the mapping of abstract system components to classic learning theory concerns, providing a "Rosetta Stone" for researchers to navigate different levels of abstraction.

Mapping of ALST levels and terms to specific learning theoretic concerns, such as capacity, convergence, and sample complexity.

As illustrated in the mapping above, different levels of the system hierarchy correspond to different theoretical concerns:

  • Elementary Level concerns map to Hardness and Distribution.
  • Cascade Level concerns map to Capacity (e.g., VC dimension), Falsifiability, and Sample Complexity.
  • Goal-Seeking Level concerns map to Convergence and Algorithmic Complexity.

This framework allows engineers to "scope" their models horizontally to see how an AGI interacts with other systems, or vertically to dive deeper into the mechanics of its goal-seeking behavior.

Conclusion

The primary result of this work is the formalization of a Learning System $S$ as a relation across eight component sets: Algorithms ($A$), Data ($D$), Parameters ($\Theta$), Consistency Relations ($G, E$), Hypotheses ($H$), Inputs ($\mathcal{X}$), and Outputs ($\mathcal{Y}$). This Mesarovician approach calls back to the original goals of AI research from the 1950s—focusing on the thinking participant directly rather than just the specific math of the problem being solved.

While the theory is highly abstract and does not prescribe specific implementation details for modern architectures like transformers, it provides a vital "latticework" for structuring design decisions in complex, large-scale AI systems. Future work will explore heterarchical relationships between learning systems and the development of practical tools for integrating ALST into Model-Based Systems Engineering (MBSE) pipelines.

Summary updated

Abstract

The solution methods used to realize artificial general intelligence (AGI) may not contain the formalism needed to adequately model and characterize AGI. In particular, current approaches to learning hold notions of problem domain and problem task as fundamental precepts, but it is hardly apparent that an AGI encountered in the wild will be discernable into a set of domain-task pairings. Nor is it apparent that the outcomes of AGI in a system can be well expressed in terms of domain and task, or as consequences thereof. Thus, there is both a practical and theoretical use for meta-theories of learning which do not express themselves explicitly in terms of solution methods. General systems theory offers such a meta-theory. Herein, Mesarovician abstract systems theory is used as a super-structure for learning. Abstract learning systems are formulated. Subsequent elaboration stratifies the assumptions of learning systems into a hierarchy and considers the hierarchy such stratification projects onto learning theory. The presented Mesarovician abstract learning systems theory calls back to the founding motivations of artificial intelligence research by focusing on the thinking participants directly, in this case, learning systems, in contrast to the contemporary focus on the problems thinking participants solve.
BibTeX
@inproceedings{cody2021mesarovician,
  author = {Tyler Cody},
  title = {{Mesarovician Abstract Learning Systems}},
  booktitle = {AGI},
  year = {2021},
  url = {https://openreview.net/forum?id=XLVLAZGml2},
}
Figure