PaperlayerView exact revision on arXiv

Source-linked AI summary

Thinking Fast and Slow in AI

Grady Booch, Francesco Fabiano, Lior Horesh, Kiran Kate, Jon Lenchner, Nick Linck, Andrea Loreggia, Keerthiram Murugesan, Nicholas Mattei, Francesca Rossi, Biplav Srivastava

arXiv:2010.06002v2cs.AI

TL;DR

AI systems remain limited in capabilities such as generalizability, common sense, and causal reasoning. This paper analyzes cognitive theories of human decision making and proposes translating insights about the causes of those capabilities into AI. It offers a high-level research direction and questions for the AI community to investigate.

Abstract

from arXiv · show

This paper proposes a research direction to advance AI which draws inspiration from cognitive theories of human decision making. The premise is that if we gain insights about the causes of some human capabilities that are still lacking in AI (for instance, adaptability, generalizability, common sense, and causal reasoning), we may obtain similar capabilities in an AI system by embedding these causal components. We hope that the high-level description of our vision included in this paper, as well as the several research questions that we propose to consider, can stimulate the AI research community to define, try and evaluate new methodologies, frameworks, and evaluation metrics, in the spirit of achieving a better understanding of both human and machine intelligence.

Motivation and Overall Vision

The paper argues that AI remains limited to narrow capabilities and proposes drawing on cognitive theories of human reasoning and decision making to investigate how human capabilities arise. It presents this as a high-level research direction, including questions intended to stimulate AI research.

Motivation and Overall Vision

  • State-of-the-art AI still lacks capabilities such as generalizability, robustness, causal analysis, abstraction, and common-sense reasoning.The paper contrasts these gaps with AI's success in narrow tasks and its reliance on large datasets and computational power.Source: Motivation and Overall Vision · p. 1
  • The authors favor integrating machine learning with symbolic and logic-based AI over relying solely on end-to-end neural networks.They cite related proposals and neuro-symbolic systems as support for the integration route.Source: Motivation and Overall Vision · p. 1
  • The paper proposes using cognitive theories of human reasoning and decision making to investigate the causes of capabilities that AI lacks.The goal is to use those theories as a source of inspiration for research questions.Source: Motivation and Overall Vision · p. 1
  • The authors analyze theories, especially Kahneman’s, and aim to connect them into a unified account of roots of human capabilities before translating them into AI.They offer a high-level overview, pose research questions, and call for a community to assess the approach.Source: Motivation and Overall Vision · p. 1

of Human Decision Making

The paper describes human decision making as cooperation between fast, intuitive processing and slower, more deliberate reasoning. Their interaction supports everyday responses as well as complex problem solving, abstraction, and adaptation.

Thinking Fast and Slow

  • System 1 handles fast, intuitive, often unconscious decisions, while system 2 handles complex situations requiring logical reasoning.The paper treats the two systems as broad classes of information processing and metaphors, not conventional multi-agent systems.Source: of Human Decision Making · pp. 1–2
  • System 1 supplies quick answers and approximate causal world models, but shortcuts and unconscious bias can make its answers wrong.The paper says people use system 1 about 95% of the time when making decisions.Source: of Human Decision Making · p. 2
  • When a problem exceeds system 1’s capacity, system 2 uses additional resources, full attention, and logical reasoning; recognizing that threshold involves introspection.Complex arithmetic and multi-criteria optimization are examples of system 2 problems.Source: of Human Decision Making · p. 2
  • Some initially difficult problems shift from system 2 to system 1 as system 2 solutions accumulate examples, but this transfer does not occur for every task.Reading in one's native language is an example of transfer; complex arithmetic is given as a counterexample.Source: of Human Decision Making · p. 2
  • System 2 can use system 1 heuristics to focus search, while system 1’s basic causal reasoning supports system 2’s more complex reasoning.The heuristics can help keep search within manageable time and space.Source: of Human Decision Making · p. 2

Other Theories

  • The paper connects abstraction and flexible collaboration to high-level stories, and links complex problem solving to both I-consciousness and M-consciousness.I-consciousness concerns recognizing processing steps for a problem; M-consciousness concerns approximate models of minds and intentions. Sophisticated I-consciousness is described as relying on M-consciousness.Source: of Human Decision Making · pp. 2–3

AI Thinking, Fast and Slow

The section treats human cognitive competencies as inspiration for AI’s missing capabilities and develops research questions for translating that insight. It focuses on system pairing, introspective governance, divergence, abstraction, and multi-agent architecture.

AI Thinking, Fast and Slow

  • The paper frames human competencies such as abstraction, generalization, causal analysis, and planning as clues for research on AI capabilities that remain lacking.It aims to stimulate the research community to study these questions and find appropriate answers.Source: AI Thinking, Fast and Slow · p. 3

System 1 and System 2 in AI

  • System 1 is loosely associated with data-driven ML and System 2 with symbolic reasoning; their symbiosis motivates hybrids and raises questions about capability boundaries and parallel reasoning.The paper asks whether machine System 2 reasoning should remain sequential or use parallel threads.Source: System 1 and System 2 in AI · p. 3
  • Hybrid systems need evaluation measures that account for ML performance scores and symbolic-reasoning correctness, but selecting those measures remains an open question.The paper asks whether metrics should vary by task and combination approach.Source: System 1 and System 2 in AI · p. 3

Introspection and Governance

  • The proposed introspection model assesses problem difficulty, risk, reward, and component competence to select procedures and govern switching between System 1 and System 2.The paper asks how introspection relates to autonomy and teaming, and what factors should trigger a switch.Source: Introspection and Governance · pp. 3–4

Divergence Resolution and Ethical Reasoning

  • For competing policies, System 2 can compare pros, cons, and counterfactual consequences, using System 1 heuristics to reduce search; the paper extends this concern to ethical reasoning.It asks which capabilities support moral judgment and how conflicting ethical theories should be modeled and deployed.Source: Divergence Resolution and Ethical… · p. 4

Abstraction, Generalization, and Knowledge

  • Adaptation is framed as abstracting and generalizing skills, then specializing them in a new environment using recognized similarities and differences to guide what to forget.The paper asks how attention can guide abstraction from raw data and what knowledge should be retained at different levels.Source: Abstraction, Generalization, and… · p. 4

Multi-agent Environment, Epistemic Reasoning, and Architectural Support

  • The paper proposes asynchronous, skill-focused agents governed and combined flexibly, and asks how epistemic reasoning and planning can support models of worlds and other agents.It also raises the broader question of which architectural choices best support this vision.Source: Multi-agent Environment, Epistemic… · p. 4
Loading 2010.06002v2…