Active Inference: First International Workshop, IWAI 2020, Co-located with ECML/PKDD 2020, Ghent, Belgium, September 14, 2020, Proceedings

Preface
Organization
Contents
Active Inference and Continuous Control
On the Relationship Between Active Inference and Control as Inference
1 Introduction
2 Formalism
3 Control as Inference
4 Active Inference
5 Encoding Value
6 Discussion
References
Active Inference or Control as Inference? A Unifying View
1 Introduction
2 Background
2.1 Problem Formulation
2.2 Variational Inference for Latent Variable Models
3 Active Inference
3.1 Free Energy of the Future
3.2 Active Inference in Practice
4 Control as Inference
4.1 Linear Gaussian Inference and Linear Quadratic Control
5 The Unifying View: Control of the Observations
6 Conclusion
References
Active Inference for Fault Tolerant Control of Robot Manipulators with Sensory Faults
1 Introduction
2 Problem Statement
3 A Fault Tolerant Scheme Based on Active Inference
4 Simulation Results
5 Discussion and Conclusion
References
A Worked Example of Fokker-Planck-Based Active Inference
1 Introduction
2 The Fokker-Planck Equation for Dynamical Systems
3 Laplace-Encoded Free Energy and Generative Models
4 A Worked Example
5 Results
6 Discussion
References
Dynamics of a Bayesian Hyperparameter in a Markov Chain
1 Introduction
2 IID Parameter Inference
2.1 Fully Observable Markov Chain
References
Online System Identification in a Duffing Oscillator by Free Energy Minimisation
1 Introduction
2 System
3 Identification
3.1 Free Energy Minimisation
3.2 Factor Graphs and Message Passing
4 Experiment
4.1 1-Step Ahead Prediction Error
4.2 Simulation Error
5 Discussion
5.1 Related Work
6 Conclusion
References
Hierarchical Gaussian Filtering of Sufficient Statistic Time Series for Active Inference
1 Introduction
2 Bayesian Inference Reduced to Mean-Tracking
2.1 Mean Tracking and Exponential Weighting
2.2 A Conjugate Prior Which Reduces Bayesian Inference to Mean Tracking for Exponential Families
3 Predictive Distributions
4 Filtering of Sufficient Statistics for Non-stationary Input Distributions
5 Discussion
References
Active Inference and Machine Learning
Deep Active Inference for Partially Observable MDPs
1 Introduction
2 Deep Active Inference Model
3 Experimental Setup
4 Results
5 Conclusion
References
Sleep: Model Reduction in Deep Active Inference
1 Introduction
2 Deep Active Inference
3 Latent Space Dimensionality Reduction and Sleep
4 Experimental Setup
5 Results
6 Conclusion
References
A Deep Active Inference Model of the Rubber-Hand Illusion
1 Introduction
2 Deep Active Inference Model
2.1 Generative Model Learning
2.2 Modelling Visuo-Tactile Stimulation Synchrony
3 Experimental Setup
4 Results
5 Conclusion
References
You Only Look as Much as You Have To
1 Introduction
2 Active Inference
3 Environment and Approach
4 Experiments
4.1 Exploring Behaviour
4.2 Goal Seeking Behaviour
5 Conclusion
A The Generative Model
References
Modulation of Viability Signals for Self-regulatory Control
1 Introduction
2 Preliminaries
2.1 Model-Free Surprisal Minimization
2.2 Expected Free Energy
3 Adaptive Control via Self-regulation
3.1 Case Study
3.2 Evaluation
4 Discussion
A Expected Free Energy with Measurements v
B Novelty and salience
C Implementation
D Drive decomposition
References
End-Effect Exploration Drive for Effective Motor Learning
1 Introduction
2 Method
2.1 A Probabilistic View to Motor Supervision
2.2 A Uniform Exploration Drive
2.3 Link with Variational Inference
3 Results
4 Conclusions
References
Learning Where to Park
1 Introduction
2 Problem Statement
3 Model Specification
3.1 The Physical Model
3.2 The Target Model
4 Experimental Validation
4.1 Setup
4.2 Results
5 Related Work
6 Conclusions
References
Active Inference: Theory and Biology
Integrated World Modeling Theory (IWMT) Implemented: Towards Reverse Engineering Consciousness with the Free Energy Principle and Active Inference
1 Integrated World Modeling Theory (IWMT) Summarized: Combining the Free Energy Principle and Active Inference (FEP-AI) Framework with Integrated Information Theory (IIT) and Global Neuronal Workspace Theory (GNWT)
2 Integrated World Modeling Theory (IWMT) Implemented
2.1 Mechanisms of Predictive Processing: Folded Variational Autoencoders (VAEs) and Self-organizing Harmonic Modes (SOHMs)
2.2 A Model of Episodic Memory and Imagination
2.3 Brains as Hybrid Machine Learning Architectures
2.4 Conclusion: Functions of Basic Phenomenal Consciousness?
3 Appendices
3.1 Appendix 1: A Model of Goal-Oriented Behavior with Hippocampal Orchestration
3.2 Appendix 2: The VAE-GAN Brain?
References
Confirmatory Evidence that Healthy Individuals Can Adaptively Adjust Prior Expectations and Interoceptive Precision Estimates
1 Introduction
2 Methods
3 Results
4 Discussion
References
Visual Search as Active Inference
1 Introduction
2 Problem Statement: Formalizing Visual Search as Accuracy Seeking
2.1 Visual Search Task
2.2 Central Processing
2.3 Accuracy Map
3 Principles: Supervised Learning of Action Selection
3.1 Peripheral Visual Processing
3.2 Motor Control
3.3 Higher Level Inference: Choosing the Processing Pathway
3.4 Learning the Accuracy Map
4 Results
5 Discussion and Perspectives
References
Sophisticated Affective Inference: Simulating Anticipatory Affective Dynamics of Imagining Future Events
1 Introduction
2 Methods
3 Results
References
Causal Blankets: Theory and Algorithmic Framework
1 Introduction
1.1 Markov Blankets
1.2 Computational Mechanics, Causal States, and Epsilon-Machines
1.3 Contribution
2 Causal Blankets as Informational Boundaries
3 Integrated Information Transcends the Blankets
4 Conclusion
A Proofs
References
Author Index