Abstract
While end-to-end learning systems are rapidly gaining capabilities and popularity, the increasing computational demands for deploying such systems, along with a lack of flexibility, adaptability, explainability, reasoning and verification capabilities, require new types of architectures. Here we introduce a classification of hybrid systems which, based on an analysis of human knowledge and intelligence, combines neural learning with various types of knowledge and knowledge sources. We present the Thrill-K architecture as a prototypical solution for integrating instantaneous knowledge, standby knowledge and external knowledge sources in a framework capable of inference, learning and intelligent control.
Introduction
Modern Deep Learning (DL) has achieved remarkable success in tasks requiring "System 1" thinking—fast, intuitive pattern recognition and correlation. However, these systems fundamentally lack "System 2" capabilities, such as logical planning, causal reasoning, and abstraction. As AI development continues, it faces three significant barriers:
- Efficiency: The computational power required for state-of-the-art models is doubling every few months, creating an unsustainable economic and environmental trajectory.
- Robustness: Current models are often brittle when encountering "unknown unknowns" outside their training distribution.
- Cognition: DL lacks the mechanisms for context and explainability—the ability to understand the "why" behind a prediction.
To overcome these hurdles, a shift from pure neural scaling to hybrid architectures that integrate structured symbolic knowledge is required.
Classification of AI Systems
The authors propose a taxonomy for AI systems based on their relationship with information and knowledge structure. This classification moves from fully internal representations to deeply structured, externalized knowledge bases.
- Class 1 (Fully Encapsulated): Standard neural models where all knowledge is implicitly stored within the network's weights.
- Class 2 (Semi-Structured Adjacent): Systems that utilize retrieval mechanisms to access external, often unstructured or semi-structured, information repositories (e.g., RAG systems).
- Class 3 (Deeply Structured Knowledge): Systems that interact with an explicit, multi-dimensional knowledge base (KB) to perform reasoned extraction. This class represents the target for the Thrill-K architecture.
The Thrill-K Architecture
The Thrill-K Architecture (Three Levels of Knowledge) provides a blueprint for Cognitive AI by partitioning knowledge into three functional layers: instantaneous, standby, and retrieved external knowledge.
In this framework, Instantaneous Knowledge resides in the neural network's parametric memory for rapid responses. Standby Knowledge is stored in an adjacent Deep Knowledge Base (often a Knowledge Graph) containing structured facts and causal models. Finally, Retrieved External Knowledge is pulled from massive digital repositories when specific data is missing internally. A central innovation is the Reasoned Extraction block, which acts as a mediator, allowing the neural network to dynamically query the structured KB during inference.
Dimensions of Knowledge and Understanding
True understanding requires more than a flat list of facts. The Thrill-K framework identifies six dimensions of knowledge necessary for cognitive intelligence: descriptive knowledge (taxonomies), models of the world (causal/physical), stories and scripts (social routines), context and source attribution, values and priorities (ethics), and unified concept references. By offloading "standby" knowledge to a structured KB, Thrill-K systems can potentially reduce neural parameter counts by orders of magnitude while improving explainability through traceable symbolic relations.
Conclusion and Future Directions
The Thrill-K architecture offers a path toward sustainable and explainable AI by integrating gradient-descent-based learning with symbolic reasoning. While theoretically robust, implementation faces significant challenges in knowledge engineering—specifically the creation of KBs that capture the breadth of human experience—and the development of hybrid algorithms that can seamlessly bridge statistical and symbolic logic. Future work must focus on standardizing these reasoning interfaces to enable widespread adoption of Class 3 systems.
Summary updated