Humans have always dreamed of automating laborious physical and intellectual tasks, but the latter has proved more elusive than naively suspected. Seven decades of systematic study of Artificial Intelligence have witnessed cycles of hubris and despair. The successful realization of General Intelligence (evidenced by the kind of cross-domain flexibility enjoyed by humans) will spawn an industry worth billions and transform the range of viable automation tasks.The recent notable successes of Machine Learning has lead to conjecture that it might be the appropriate technology for delivering General Intelligence. In this book, we argue that the framework of machine learning is fundamentally at odds with any reasonable notion of intelligence and that essential insights from previous decades of AI research are being forgotten. We claim that a fundamental change in perspective is required, mirroring that which took place in the philosophy of science in the mid 20th century.We propose a framework for General Intelligence, together with a reference architecture that emphasizes the need for anytime bounded rationality and a situated denotational semantics. We given necessary emphasis to compositional reasoning, with the required compositionality being provided via principled symbolic-numeric inference mechanisms based on universal constructions from category theory.
• Details the pragmatic requirements for real-world General Intelligence.
• Describes how machine learning fails to meet these requirements.
• Provides a philosophical basis for the proposed approach.
• Provides mathematical detail for a reference architecture.
• Describes a research program intended to address issues of concern in contemporary AI.
The book includes an extensive bibliography, with ~400 entries covering the history of AI and many related areas of computer science and mathematics.The target audience is the entire gamut of Artificial Intelligence/Machine Learning researchers and industrial practitioners. There are a mixture of descriptive and rigorous sections, according to the nature of the topic. Undergraduate mathematics is in general sufficient. Familiarity with category theory is advantageous for a complete understanding of the more advanced sections, but these may be skipped by the reader who desires an overall picture of the essential concepts
Author(s): Jerry Swan, Eric Nivel, Neel Kant, Jules Hedges, Timothy Atkinson, Bas Steunebrink
Series: Studies In Computational Intelligence | 1049
Edition: 1
Publisher: Springer
Year: 2022
Language: English
Commentary: TruePDF
Pages: 142
Tags: Computational Intelligence; Artificial Intelligence; Data Engineering
Foreword by Melanie Mitchell
Foreword by David Spivak
Contents
1 Introduction
Part I Requirements
2 Background
2.1 What we Mean by General Intelligence
2.2 Science as Extended Mind
2.3 The Death of `Good Old-Fashioned AI'
3 Where is My Mind?
3.1 A Sanity Check
3.2 Real-World Machine Learning
4 Challenges for Deep Learning
4.1 Compositionality
4.2 Strong Typing
4.3 Reflection
4.4 Implications and Summary
5 Challenges for Reinforcement Learning
5.1 A Priori Reward Specification
5.2 Sampling: Safety and Efficiency
6 Work on Command: The Case for Generality
6.1 Goals and Constraints
6.2 Planning
6.3 Anytime Operation
Part II Semantically Closed Learning
7 Philosophy
7.1 The Problem of Machine Induction
7.2 Semantically Closed Learning (SCL)
7.3 Baseline Properties of SCL
7.4 High-Level Inference Mechanisms of SCL
7.5 Intrinsic Motivation and Unsupervised Learning
8 Architecture
8.1 SCL as a Distributed/Localist Hybrid
8.2 Reference Architecture
9 A Compositional Framework
9.1 Categorical Cybernetics
9.2 Hypothesis Generation
9.3 Abstraction and Analogy
9.4 Abduction
10 2nd Order Automation Engineering
10.1 Behavioral System Engineering
10.2 Reactive Synthesis
10.3 Proactive Synthesis
10.4 Safety
11 Prospects
11.1 Summary
11.2 Research Topics
11.2.1 Choice of Expression Language
11.2.2 Compositional Primitives
11.2.3 Links with Behavioral Control
11.2.4 Pragmatics via `Causal Garbage Collection'
11.3 Conclusion
Appendix Bibliography
