TL;DR
This paper argues that "ineffable" cognitive phenomena like intuition, insight, and inspiration can be explained and modeled by artificial intelligence. By defining these concepts operationally and demonstrating their replication in computer programs, the authors provide empirical evidence that these are not mystical processes but rather emergent properties of complex information processing.
What Problem Does This Paper Address?
For decades, certain aspects of human cognition – intuition, insight, and inspiration – have been considered mysterious and beyond the reach of scientific explanation, often relegated to philosophical or even mystical realms. This paper tackles the challenge of demystifying these phenomena by proposing that they are not inherently inexplicable but can be understood and modeled as outcomes of information processing. The authors aim to bridge the gap between the subjective experience of these cognitive states and objective, empirical scientific inquiry, particularly through the lens of artificial intelligence and cognitive psychology.
Methods
The core methodology involves defining cognitive phenomena operationally, meaning establishing clear, testable criteria for when they occur. For instance, intuition is defined by rapid problem-solving with an inability to articulate the steps taken. Insight is characterized by a sudden shift in problem representation after a period of unsuccessful work, often accompanied by a feeling of understanding. Inspiration is framed as the generation of novelty.
Once defined, these phenomena are then modeled using computer programs. The paper highlights existing AI systems like EPAM (for recognition and rote learning) and the General Problem Solver (GPS) or Soar (for heuristic search). The authors propose that combining these mechanisms can simulate intuition and insight. For example, EPAM's recognition capabilities can explain rapid, seemingly intuitive diagnoses, while the interrupt and redirection mechanisms in a GPS-like system can model insight. The BACON system is presented as a model for scientific discovery, demonstrating how "generate and test" mechanisms, guided by data, can lead to the discovery of scientific laws and even new concepts.
Datasets
Not discussed in the paper. The paper focuses on theoretical modeling and empirical evidence from existing psychological studies and experiments, rather than introducing new datasets for training AI models.
Key Insights and Ideas
- Operational Definitions are Key: The paper strongly emphasizes that to scientifically study phenomena like intuition, they must first be defined by observable, testable criteria. This moves them from the realm of the ineffable to the empirically investigable.
- Intuition as Recognition: Intuition is reframed as a form of rapid recognition based on extensive prior experience. This explains why experts often exhibit intuition in their domain but not outside it, and why it can be fallible.
- Insight as a Shift in Representation: Insight is explained as a consequence of a system (human or computer) pausing an unsuccessful search, shifting its attention to a new problem representation, and then resuming the search with a more productive approach.
- Discovery as Combinatorics and Heuristics: Scientific discovery, including the generation of novelty and new concepts, is presented not as a divine spark but as a result of combinatorial processes ("generate and test") guided by selective search heuristics, as exemplified by the BACON system.
- AI as a Theory of Cognition: Computer programs are not just tools but can serve as precise "languages of theory" for cognitive psychology, analogous to how differential equations are used in physics. Testing these programs against human behavior provides empirical validation.
- Philosophy's Role: The authors argue that philosophical speculation is insufficient for evaluating cognitive theories; empirical evidence is paramount.
Key Results
The paper presents a theoretical framework and empirical evidence supporting the AI modeling of intuition, insight, and inspiration.
- EPAM's Recognition: EPAM, when presented with familiar stimuli, can achieve recognition in a fraction of a second, mimicking the speed of human intuition. It also demonstrates the inability to report the internal steps of this recognition, aligning with the definition of intuition.
- Mutilated Checkerboard Problem: Experiments with this problem showed that subjects who solved it (often after a hint or a period of incubation) exhibited behaviors consistent with the proposed model of insight, involving a shift in representation and rapid solution.
- BACON's Discoveries: The BACON system successfully rediscovered several important scientific laws (e.g., Kepler's Third Law, Ohm's Law) from historical data, demonstrating that scientific discovery can arise from systematic, data-driven search processes. It also showed the ability to reinvent concepts like "inertial mass."
- Incubation Explained: Incubation is explained through mechanisms like forgetting unproductive search paths in short-term memory and redirection of attention, rather than requiring "unconscious work."
Limitations and Future Work
- Empirical Validation: While the paper presents strong theoretical arguments and cites empirical evidence, the authors acknowledge that the "insight" model is presented more tentatively than the "intuition" model, suggesting a need for further validation.
- Scope of AI Models: The paper focuses on specific aspects of discovery and cognition. While it argues that these phenomena are explainable by current AI mechanisms, the full breadth of human creativity and discovery likely requires further exploration and integration of various AI approaches.
- Physiological Foundations: The authors explicitly state they have said "almost nothing" about the brain's physiological underpinnings, focusing solely on the information-processing level. They acknowledge that understanding the links between physiological mechanisms and symbolic processes is a crucial area for future work.
- Complexity of Human Cognition: The paper acknowledges that while AI can simulate these phenomena, the complexity of human cognition, especially in ill-structured tasks, still presents significant challenges. The authors invite further empirical study to refine and extend these models.
Summary updated