Thinking Fast and Slow in AI: the Role of Metacognition (2021)

Reference ID: arXiv:2110.01834
Submission Date: October 5, 2021
Category: Computer Science $\rightarrow$ Artificial Intelligence (cs.AI)

✍️ Authorship

The research was conducted by a collaborative team including:

• Marianna Bergamaschi Ganapini
• Murray Campbell
• Francesco Fabiano
• Lior Horesh
• Jon Lenchner
• Andrea Loreggia
• Nicholas Mattei
• Francesca Rossi
• Biplav Srivastava
• Kristen Brent Venable

---

🔍 Executive Summary

While the field of Artificial Intelligence has experienced an explosion of growth, leading to ubiquitous daily applications, the current state of the art is primarily characterized by General Intelligence Narrow AI. These systems are designed for highly specific goals, such as:

• Natural Language Processing (NLP)
• Image interpretation
• Predictive analytics
• Data classification

The success of these tools is often a function of the following relationship: $\text{AI Performance} \approx f(\text{Algorithmic Efficiency}, \text{Massive Datasets}, \text{Computational Power})$

The Intelligence Gap

Despite these leaps, AI still lacks the holistic capabilities inherent in human cognition. The authors identify several missing links:

• Ability to reason across diverse domains
• Metacognitive awareness
• Flexible problem-solving beyond training data

"We argue that a better study of the mechanisms that allow humans to have these capabilities can help us understand how to imbue AI systems with these competencies."

---

🧠 Proposed Framework: The Dual-Process Architecture

The authors draw inspiration from Daniel Kahneman's psychological theory of Thinking, Fast and Slow. They propose a multi-agent architecture that bifurcates problem-solving into two distinct modes:

Conceptual Logic Flow

def solve_problem(incoming_problem):
    if system_1.can_handle(incoming_problem):
        return system_1.react(past_experience)
    else:
        # Activate metacognition to trigger deliberative reasoning
        return system_2.reason_and_search(optimal_solution)

System Architecture Visualization

---

🛠️ Supporting Knowledge Bases

To function effectively, both agent types rely on two critical internal models:

1. The World Model: A repository of domain-specific knowledge regarding the external environment.
2. The Self Model: A metacognitive layer containing:
   • Records of past actions taken by the system.
   • An inventory of the solvers' specific skills.

---

Additional Metadata:

• DOI: 10.48550/arXiv.2110.01834
• Source: Provided via arXivLabs, an open framework for community-driven AI research features.