Abstract
This paper introduces a new metamodel-based knowledge representation that significantly improves autonomous learning and adaptation. While interest in hybrid machine learning / symbolic AI systems leveraging, for example, reasoning and knowledge graphs, is gaining popularity, we find there remains a need for both a clear definition of knowledge and a metamodel to guide the creation and manipulation of knowledge. Some of the benefits of the metamodel we introduce in this paper include a solution to the symbol grounding problem, cumulative learning, and federated learning. We have applied the metamodel to problems ranging from time series analysis, computer vision, and natural language understanding and have found that the metamodel enables a wide variety of learning mechanisms ranging from machine learning, to graph network analysis and learning by reasoning engines to interoperate in a highly synergistic way. Our metamodel-based projects have consistently exhibited unprecedented accuracy, performance, and ability to generalize. This paper is inspired by the state-of-the-art approaches to AGI, recent AGI-aspiring work, the granular computing community, as well as Alfred Korzybski's general semantics. One surprising consequence of the metamodel is that it not only enables a new level of autonomous learning and optimal functioning for machine intelligences, but may also shed light on a path to better understanding how to improve human cognition.
Introduction
Modern artificial intelligence faces a "Goldilocks" problem. Subsymbolic approaches, such as Deep Learning, excel at pattern recognition but remain brittle, requiring massive labeled datasets and struggling with generalization. Conversely, symbolic AI (reasoning) can generalize and provide explainable logic but often suffers from "combinatorial explosion," where the complexity of relationships becomes computationally insurmountable in real-time.
Furthermore, the "symbol grounding problem" persists: how does a machine bridge the gap between raw pixels (subsymbolic) and abstract concepts like "shelf" or "product" (symbolic)? This paper presents the Deep Fusion Reasoning Engine (DFRE), a neurosymbolic framework that addresses these issues by organizing knowledge into a strict hierarchy of abstraction inspired by Alfred Korzybski’s General Semantics.
The DFRE Metamodel
The core of the framework is a hierarchical knowledge representation that structures information into four distinct levels of abstraction, moving from raw sensory input to high-level meta-goals.
- L0 (Physical World): Raw sensor data, such as video frames or time-series signals.
- L1 (Spatial/Relational): Initial knowledge graphs representing spatial relations (e.g., "Rectangle A is inside Rectangle B").
- L2 (Abstract Reasoning): High-level concepts (e.g., "Shelf," "Product") and statistical learning.
- L* (Meta-Goals): Self-monitoring, motivations, and self-repair mechanisms.
By strictly separating these levels, the system avoids "knowledge corruption"—the mixing of raw data with abstract concepts—which often leads to brittleness in traditional models.
Technical Architecture and Implementation
The DFRE system integrates several advanced components to enable real-time reasoning under uncertainty. Central to this is the OpenNARS reasoning engine, which operates under the Assumption of Insufficient Knowledge and Resources (AIKR). This allows the system to make logical inferences even when data is incomplete.
To manage the complexity of visual data, the system employs a Focus of Attention (FoA) mechanism. Much like human vision, FoA segments the environment into manageable "covers," processing them sequentially and merging results. This prevents the reasoning engine from being overwhelmed by the thousands of potential relations generated by raw sensor input.
Experimental Evaluation: Retail Object Recognition
The framework was evaluated in a retail store use case focused on unsupervised object class recognition. Unlike traditional deep learning, the system required no training set or labeled data; it learned directly from the environment using expert-provided rules at the L2 level.
The impact of the Focus of Attention (FoA) mechanism was particularly significant, as shown in the table below:
| Metric | Without FoA | With FoA |
|---|---|---|
| Product Precision | 80.70% | 96.36% |
| Shelf Precision | 8.82% | 82.35% |
| Overall Accuracy | 46.30% | 94.73% |
Without FoA, the reasoning engine was overwhelmed by noise (1,478 premises from just 152 rectangles). With FoA, the system successfully filtered this noise, achieving near-perfect accuracy in identifying products and a 73% improvement in shelf identification.
Conclusion
The DFRE metamodel demonstrates that by structuring knowledge according to levels of abstraction and employing attention-based filtering, AI systems can achieve high accuracy and generalization without the need for massive labeled datasets. Future work will explore "QA-Cooperative networks," where multiple agents communicate to describe unseen environments, further testing the framework's ability to transfer knowledge across different modalities.
Summary updated