Artificial General Intelligence: 16th International Conference, AGI 2023, Stockholm, Sweden, June 16–19, 2023, Proceedings

Description

This book constitutes the refereed proceedings of the 16th International Conference on Artificial General Intelligence, AGI 2023, held in Stockholm, Sweden in June 2023. The 35 full papers and one short paper presented in this book were carefully reviewed and selected from 72 submissions. The papers cover topics from foundations of AGI, to AGI approaches and AGI ethics, to the roles of systems biology, goal generation, and learning systems, and so much more. Preface Organization Contents On VEI, AGI Pyramid, and Energy 1 Introduction 2 Intelligence 2.1 Intelligence and Power Consumption – Part I: The Natural Trend 2.2 Intelligence and Power Consumption – Part II: Human Brain 2.3 Intelligence, NPS and Time 3 Software vs. Hardware 3.1 Software vs. Hardware – Part I: Computational Power 3.2 Software vs. Hardware – Part II: Hardware Architecture 3.3 Software vs. Hardware – Part III: Singularity 4 The Roles of the AGI Society References Elements of Cognition for General Intelligence 1 Introduction 1.1 How Does Cognition Work? 2 Reflex Systems 3 The Spatial Basis of Cognition 4 Sensorimotor Encoding of Affordances and Outcomes 5 Top Level Control 6 Discussion References Comparing NARS and Reinforcement Learning: An Analysis of ONA and Q-Learning Algorithms 1 Introduction 2 Methods 2.1 RL and Tabular Q-Learning 2.2 NARS and ONA 3 Setups and Environments 4 Results and Discussion 5 Conclusion and Future Work References On the Computation of Meaning, Language Models and Incomprehensible Horrors 1 Introduction 1.1 Grice's Foundational Theory of Meaning 1.2 A Foundational Theory of Foundational Theories 2 Meaning, From the Top Down 2.1 Extending the Formalism 2.2 Interpretation 3 Communication of Meaning 3.1 Ascribing Intent 3.2 From Manipulation to Meaningful Communication 4 Talking to a Machine References The Optimal Choice of Hypothesis Is the Weakest, Not the Shortest 1 Introduction 2 Background Definitions 3 Formalising Induction 4 Proofs 5 Experiments 5.1 Setup 5.2 Trials 5.3 Results 6 Concluding Remarks References Emergent Causality and the Foundation of Consciousness 1 Introduction 2 Additional Background 2.1 List of Definitions 2.2 Premises 3 Emergent Causality 3.1 The do Operator as a Variable in Disguise 3.2 Emergent Representation of Interventions 3.3 When Will Induction Distinguish Intervention From Observation? 4 Awareness 4.1 Intent 4.2 How Might We Represent the Mirror Symbol Hypothesis? 4.3 Consciousness References The M Cognitive Meta-architecture as Touchstone for Standard Modeling of AGI-Level Minds 1 Arithmetic as the Initial Target 2 The Formal Procedure, for Arithmetic Cognition 2.1 Peano Arithmetic to Anchor Arithmetic Cognition 2.2 Definition of the Procedure 3 The M Cognitive Meta-architecture: Key Attributes 4 Applying the Procedure 4.1 Exploration of NARS 4.2 Exploration of SNePS and GLAIR 5 Related Work 6 Objections 6.1 ``But What About Purely Numerical Approaches to AGI?'' 6.2 ``Math is Merely Manufactured'' 7 Conclusion and Next Steps References Causal Reasoning over Probabilistic Uncertainty 1 Introduction 2 Related Work 3 Methodology 3.1 Autocatalytic Endogenous Reflective Architecture - AERA 3.2 Background Assumptions 3.3 Modeling of Probabilistic Uncertainty 4 Reasoning Algorithm 5 Discussion and Future Work References Probabilistic Logic Networks for Temporal and Procedural Reasoning 1 Introduction 2 Probabilistic Logic Networks Recall 2.1 Elementary Notions 2.2 Inference Rules 3 Temporal Probabilistic Logic Networks 3.1 Temporal Operators 3.2 Temporal Rules 3.3 Example 4 Procedural Reasoning References Rational OpenCog Controlled Agent 1 Introduction 2 OpenCog Framework Recall 3 Rational OpenCog Controlled Agent 3.1 Memory 3.2 Processes 3.3 Control 3.4 Learning 4 Experiment in a Simple Minecraft Environment 5 Conclusion References Towards Cognitive Bots: Architectural Research Challenges 1 Introduction 2 The Transduction Problem 3 The Behavior Problem 4 Bot Communication Challenges 5 Integration and Cooperation Challenges 6 Bot Ethics and Trust 7 Conclusion References Bridging AGI Theory and Practice with Galois Connections 1 Introduction 2 Discrete Decision Systems 3 Combinatory-Operation-Based Function Optimization 4 Cognitive Processes as COFO-Guided Metagraph Transformations 5 COFO Processes as Galois Connections 5.1 Greedy Optimization as Folding 5.2 Galois Connection Representations of Dynamic Programming Decision Systems Involving Mutually Associative Combinatory Operations 6 Associativity of Combinatory Operations Enables Representing Cognitive Operations as Folding and Unfolding 7 Challenges and Prospects References Comparative Reasoning for Intelligent Agents 1 Introduction 2 Related Work 3 Formalization of Comparative Relation 4 Experiment 5 Conclusion References Primum Non Nocere: The Ethical Beginnings of a Non-Axiomatic Reasoning System 1 Introduction 1.1 NARS in Health Care 2 Non-axiomatic Reasoning System 2.1 Term Logic 2.2 Statement Types 2.3 Arbitrary Relations 2.4 Higher Order Statements 3 Example Scenario 4 Normative Ethical Theories 4.1 Consequentialism 4.2 Deontological and Virtue Ethics 5 Conclusions References Memory System and Memory Types for Real-Time Reasoning Systems 1 Introduction 2 Related Works 3 NARS Considerations 4 Types of Memory in NARS 5 Experiments 6 Discussions 7 Conclusion References Stimulus Equivalence in NARS 1 Introduction 2 OpenNARS for Applications (ONA) 3 The Matching-to-Sample Task 4 Stimulus Equivalence 5 Matching-to-sample Task in NARS 6 Contingency Entailment 7 Acquired Relations 8 Symmetry Based on Contingency Entailment 9 Transitivity Based on Acquired Relations 10 Equivalence as Combined Symmetry and Transitivity 11 Generalizing Outside the Matching-to-Sample Task 12 Discussion References Context-Rich Evaluation of Machine Common Sense 1 Background 2 Proposed Evaluation Protocol 3 Experimental Demonstration 3.1 Context 1: Camping Trip 3.2 Context 2: Vacationing Abroad 4 Discussion References Indications of Suitable Algorithms for an AGI 1 Introduction 1.1 Background: Humans and AGI Learn from their Conversations 1.2 Purpose: To Look for Indications of Suitable Algorithms for an AGI 1.3 Lessons Learned from Previous Implementations 1.4 Method for Evaluating the Usefulness of the Examples 2 Similar Functionality in People and AGIs 2.1 Typed Variables Facilitate Reasoning on a Generic Level 2.2 The AGI Should be Able to Store Complex Patterns in Its Working Memory 2.3 The Context Should be Used When Interpreting the Users Input 2.4 The Dynamic Memory of an AGI Should be Independent of Schemas 2.5 Knowledge Should be Continuously Updated in Dialogues with Users 2.6 The System Must Always be Able to Explain How It has Been Reasoning 2.7 An AGI Should be Able to Downgrade Rarely Used Knowledge 3 More Advanced Functionality in Humans and AGIs 3.1 Knowledge Should be Rebalanced in Hierarchical Structures 3.2 A Learning System Must be Able to Generate and Test Hypotheses 3.3 An AGI Should Use Creative Inspection of Its Own Knowledge Structures 3.4 There Should be a Consistency in the Direction of the Reasoning Process 3.5 A Continuous Induction of Gestalts for Emphasizing Feelings of Context 4 Conclusions 4.1 An AGI Can Handle Knowledge in a Similar Way as Humans Do 4.2 A Generic and Flexible Representation of Knowledge is Needed References Adaptive Predictive Portfolio Management Agent 1 Introduction 2 Advanced Agent Architecture 3 Experimental Results 3.1 Multi-asset Multi-strategy Adaptive Portfolio Management 3.2 Experiential Learning Based on Simulation and Backtesting 3.3 Predictive Adaptive Market Making 4 Conclusion and Future Work References A Vertical-Horizontal Integrated Neuro-Symbolic Framework Towards Artificial General Intelligence 1 Introduction 1.1 Vertical and Horizontal Neuro-Symbolic Approaches 2 The Hybrid Framework for AGI Research 3 Implementation on CLEVR 3.1 The Agent 3.2 The Spiking Reasoning Network 4 Implementation on CLEVRER 5 Results and Related Works 6 Discussion References Rethinking the Physical Symbol Systems Hypothesis 1 Introduction 2 Physical Symbol Systems 3 Critiquing the Physical Symbol Systems Hypothesis 4 Rethinking Symbol Systems 5 Neural Networks as Hybrid Symbol Systems 6 Conclusion References On Relation Between Facial Expressions and Emotions 1 Introduction 2 Mathematical Statement of the Problem 3 Statistical Solution 4 Practical Approach 5 Conclusions References Evaluation of Pretrained Large Language Models in Embodied Planning Tasks 1 Introduction 2 Metrics to Evaluate the Generated Plan 3 Method 3.1 Problem Formulation 3.2 Subtask Evaluation Mode 3.3 Autoregressive Plan Generation 3.4 Plan Evaluation 4 Experiments 4.1 ALFRED Environment 4.2 Prompt Engineering 5 Results 6 Conclusion References Alien Versus Natural-Like Artificial General Intelligences 1 Introduction – The Need to Consider What Type of System is Producing the Intelligent Behavior 2 Definitions of Artificial Intelligence (AI) and Artificial General Intelligence (AGI) 3 A Definition of a Natural-Like Artificial General Intelligence 4 An Example of a Natural-Like Artificial General Intelligence 5 Alien-Like AGI Versus Natural-Like AGI 6 Discussion References Computing with Categories in Machine Learning 1 Introduction 1.1 Notation 2 Foundations of DisCoPyro 2.1 Model Learning and Variational Bayesian Inference 3 Example Application and Training 3.1 Deep Probabilistic Program Learning with DisCoPyro 3.2 Experimental Results and Performance Comparison 4 Discussion References ADAM: A Prototype of Hierarchical Neuro-Symbolic AGI 1 Introduction 2 Deep Control: The Architecture of Intelligence 2.1 Reinforcement Learning as a Mental Language 2.2 Deep Control: Scale-Free Thinking 3 ADAM: Implementing Intelligence 3.1 ADAM's Design 3.2 ADAM's Prospects 4 Related Work 5 Conclusion References Electronic Education Machine AGI-EEdu 1 Introduction 2 Theoretical Background 2.1 A New Science Has Been Born 2.2 About Theory of Intelligence 3 Research Questions 4 Project - ``AGI-EEdu'' Machine 4.1 AGI-EEdu's Main Task 4.2 Statistical Data from Möbius MCW 4.3 Statistical Data from LMS Canvas 4.4 Methodology of the Project 4.5 Non-Axiomatic Reasoning System (NARS) 4.6 Narsese Language 4.7 NAL5 and NAL6 - Higher Order Terms 4.8 Minimal Narsese Course Success Reasoning Example References Can Language Models Be Used in Multistep Commonsense Planning Domains? 1 Introduction 2 Benchmark Construction 3 Metrics 4 Experiments 5 Conclusion and Future Work References Explicit Goal-Driven Autonomous Self-Explanation Generation 1 Introduction 2 Related Work 3 Definitions 4 Goal-Driven Explanation Generation 5 Explanation Generation in AERA 6 Conclusion References Addressing the Unsustainability of Deep Neural Networks with Next-Gen AI 1 Introduction 2 Deep Limitations of Deep Neural Networks 2.1 DNNs: Expensive to Develop and Use 2.2 The Limited ``learning'' of Statistics-Based Systems 2.3 DNN Autonomous Learning After It Leaves the Lab: `Undefined' 2.4 DNNs and Social Inequality 2.5 DNNs' Domination of the AI Narrative 3 Summary of Limitations 4 Breaking the Stalemate Through Innovation 5 Sustainable Automation via AGI References NUTS, NARS, and Speech 1 Introduction 2 Background 3 Literature Review 4 Method 5 Results 6 Discussion References Computational-Level Analysis of Constraint Compliance for General Intelligence 1 Introduction 2 Sources of Complexity in Constraint Processing 2.1 Partial Observability 2.2 Dynamic, Fail-Hard Environments 2.3 Abstract and Poorly-Defined Constraints 2.4 Implicit Context Specification 2.5 Interactions and Conflicts Among Constraints, Tasks, and Contexts 3 Computational-Level Analysis 3.1 Functional Role 3.2 Processing Steps for Constraint Compliance 4 Exploratory Algorithmic-Level Prototype 5 Discussion and Implications References Self-Comprehension for More Coherent Language Generation 1 Introduction 2 A Simple Problem Domain 3 The Architecture 3.1 The Bubble Chamber 3.2 The Worldview and Focus 3.3 The Coderack 3.4 Satisfaction and Randomness 4 Pattern Recognition on Many Levels 5 The Program's Behaviour 5.1 An Example of the Program Running 5.2 Emergent Pipelines 5.3 Problems the Program Encounters 6 Future Work: Meta-Level Pattern Recognition 7 Conclusion References An Adaptive Vision Architecture for AGI Systems 1 Introduction 2 Architecture 3 Training and Implementation 4 Experiments and Results 5 Conclusion References A Unified Structured Framework for AGI: Bridging Cognition and Neuromorphic Computing 1 Introduction 1.1 Research Status 1.2 Motivation 2 Methodology 2.1 Framework Overview 2.2 Cross-hierarchy Computation Mechanism 2.3 Initiatory Verification 3 Discussion References Coherence in Intelligent Systems 1 Introduction 2 Hypothesis 2.1 Hypothesis 1: AGI as a Solution Searcher 2.2 Hypothesis 2: Non-equilibrium Stable States as Solution 3 Neuronal Coherence as an Universal Indicator 4 Neuronal Coherence as Most Informative States 5 Neuronal Coherence as the Outcome of Coherent Self-evidencing 5.1 Binding Problem in Perception 5.2 Model Description 5.3 Emergence of Neuronal Coherence as Binding Solution 5.4 Neuronal Coherence from Coherent Self-evidencing Loop 6 Coherence as a Measure of Intelligence? 7 Conclusion References Author Index