As a discipline, human-centered AI (HCAI) aims to create Artificial Intelligence (AI) systems that collaborate with humans, enhancing human capabilities and empowering humans to achieve their goals. That is, the focus amplify and augment rather than displace human abilities. HCAI seeks to preserve human control in a way that ensures artificial intelligence meets our needs while also operating transparently, delivering equitable outcomes, and respecting human rights and ethical standards. Design methods that enable representation of and adherence to values such as privacy protection, autonomy (human in control), and non-discrimination are core to HCAI. These are themes closely connected to some of the most fundamental challenges of AI.
Artificial neural networks provide a distributed computing technology that can be trained to approximate any computable function, and have enabled substantial advances in areas such as computer vision, robotics, speech recognition and natural language processing. This chapter provides an introduction to Artificial Neural Networks, with a review of the early history of perceptron learning. It presents a mathematical notation for multi-layer neural networks and shows how such networks can be iteratively trained by back-propagation of errors using labeled training data. It derives the back-propagation algorithm as a distributed form of gradient descent that can be scaled to train arbitrarily large networks given sufficient data and computing power.
Black-box Artificial Intelligence (AI) systems for automated decision making are often based on over (big) human data, map a user’s features into a class or a score without exposing why. This is problematic for the lack of transparency and possible biases inherited by the algorithms from human prejudices and collection artefacts hidden in the training data, leading to unfair or wrong decisions. The future of AI lies in enabling people to collaborate with machines to solve complex problems. This requires good communication, trust, clarity, and understanding, like any efficient collaboration. Explainable AI (XAI) addresses such challenges, and for years different AI communities have studied such topics, leading to different definitions, evaluation protocols, motivations, and results. This chapter provides a reasoned introduction to the work of Explainable AI to date and surveys the literature focusing on symbolic AI-related approaches. We motivate the needs of XAI in real-world and large-scale applications while presenting state-of-the-art techniques and best practices and discussing the many open challenges.
Artificial intelligence, and in particular Machine Learning methods, is fast gaining ground. Algorithms trained on large datasets and comprising numerous hidden layers, with up to a trillion parameters, are becoming common. Such models are difficult to explain to lay users with little understanding of the basis of machine learning, but they are also hard to interpret for those who designed and programmed them. The calculations that are carried out by the algorithm are not assigned an easily understandable meaning, aside from there being far too many of these calculations to actually follow. The outputs of algorithms are, as a result, hard to predict and to explain. Why did the algorithm output that there is a cat on this picture? We don’t really know, certainly not without additional help in the form of explainability tools.
The first section, Introduction to Human-centered AI, presents the main definitions and concepts covered in this volume.
The second section, Human-centered Machine Learning, includes several chapters on machine learning ranging from basic concepts of neural networks to interactive learning. This section also describes modern approaches such as transformers in natural language processing, speech processing, vision and multi-modal processing.
The third section, Explainable AI, deals with both technical and philosophical concepts. The section includes a conceptual overview of computational cognitive vision together with practical demonstrations.
The fourth section, ethics, law and society AI, introduces main concepts of Ethics and Law. This section also discusses ethics in communication.
The fifth section, Argumentation, focuses on concepts of arguments and attacks. The concepts are illustrated with several concrete examples in cognitive technologies of learning and explainable inference or decision making.
The last section, Social Simulation, deals with agent-based social simulations that are used to investigate complex phenomena within social systems. The chapters show how they could be designed, evaluated and employed by decision makers.
Author(s): Mohamed Chetouani, Virginia Dignum, Paul Lukowicz, Carles Sierra
Series: Lecture Notes in Artificial Intelligence, 13500
Publisher: Springer
Year: 2023
Language: English
Pages: 434
Preface
Organization
Contents
Introduction to Human-Centered AI
The Advanced Course on Human-Centered AI: Learning Objectives
1 Introduction
2 Specific Learning Objectives
2.1 Human-Centered Machine Learning
2.2 Explainable AI
2.3 Ethics, Law and Society
2.4 Argumentation
2.5 Social Simulation
Conceptual Foundations of Human-Centric AI
1 Introduction
2 What is the Distinction Between Reactive and Deliberative Intelligence?
3 Are Reactive and Deliberative Intelligence both Needed for AI?
4 What is Understanding?
5 What are Meanings?
6 Why is Understanding Hard?
7 Rich Models Have Multiple Layers and Perspectives
8 Human Models Often Take the Form of Narratives
9 Open Issues for Understanding
10 Conclusions
References
Human-Centered Machine Learning
Machine Learning with Neural Networks
1 Machine Learning
2 Perceptrons
3 Artificial Neural Networks
3.1 Gradient Descent
3.2 Multilayer Neural Networks
3.3 Activation Functions
3.4 Backpropagation as Distributed Gradient Descent
References
Generative Networks and the AutoEncoder
1 Generative Networks
2 The AutoEncoder
3 Background from Information Theory
3.1 Entropy
3.2 Computing Entropy
3.3 Cross Entropy
3.4 Binary Cross Entropy
3.5 Categorical Cross Entropy Loss
3.6 The Kullback-Leibler Divergence
3.7 Sparsity
4 Variational Autoencoders
5 Generative Adversarial Networks
References
Convolutional Neural Networks
1 Convolutional Neural Networks
1.1 Convolution
1.2 Early Convolutional Neural Networks: LeNet
1.3 Convolutional Network Hyper-parameters
1.4 The LeNet-5 Architecture
2 Classic CNN Architectures
2.1 AlexNet
2.2 VGG-16 - Visual Geometry Group 16 Layer Architecture
2.3 YOLO: You Only Look Once
2.4 YOLO-9000 (YOLOv2)
3 Conclusions
References
Transformers in Natural Language Processing
1 Introduction
2 Writing Machines: Language Models
2.1 The Simplest Model
2.2 Word Order
2.3 Neural Models: Smoothing the Context Space
2.4 Defining the Vocabulary
3 The Transformer Model
3.1 Attention, a Fundamental Mechanism
3.2 Causal Transformer as Pure Language Models
3.3 Transformers as Representations: Bert and Its Clones
3.4 Computational Costs of Transformer-ing
3.5 Transformers: A Flexible Architecture
4 Towards Multilingualism
4.1 Neural Machine Translation: Conditional Text Generation
4.2 Multilingual Representations, Multilingual Translations
4.3 One Model to Translate Them All
4.4 Machine Translation as a Generic Task
4.5 Summary
5 Conclusion
References
Vision and Multi-modal Transformers
1 Introduction
2 From Convolutional Neural Networks to Transformers
2.1 Convolutional Neural Networks (CNN)
2.2 Self-attention
3 Transformers for Computer Vision
3.1 Introduction of Positional Encodings
3.2 Dynamic Positional Encodings
3.3 Convolution and Transformers
4 Transformers for Multimedia Data
5 Conclusion
References
Transformers in Automatic Speech Recognition
1 Introduction
2 Historical Perspective
2.1 The Prehistory
2.2 Statistical-Based ASR
2.3 Neural-Based ASR
2.4 End-to-End Approach
3 Transformers for ASR
3.1 Attention for Speech
3.2 Attention in Speech
3.3 Transformer-Based ASR Models
4 Transformers as Representations of Speech
4.1 Self-attention in Speech
4.2 Wav2vec
5 Conclusion
References
Interactive Robot Learning: An Overview
1 Introduction
2 Tutorial Scope and Resources
2.1 Learning Objectives
2.2 Notations
2.3 Acronyms
2.4 Selected Relevant Readings
3 Background
3.1 Fundamentals of Reinforcement Learning
3.2 Robot Learning
4 Interactive Machine Learning vs. Machine Learning
4.1 Human Interventions in the Supervised Machine Learning Process
4.2 Human Interventions in the Interactive Machine Learning Process
5 Overview of Human Strategies
6 Feedback
6.1 Representation
6.2 Definition
6.3 Learning from Evaluative Feedback
6.4 Reward Shaping
6.5 Value Shaping
6.6 Policy Shaping
6.7 Example: The TAMER Architecture
6.8 Limitations
7 Demonstrations
7.1 Representation
7.2 Definition
7.3 Methods
7.4 Learning from Demonstrations
7.5 Behavioral Cloning
7.6 Imitation Learning as a Sequential Decision-Making Problem
7.7 Example: Inverse Reinforcement Learning
7.8 Limitations
8 Instructions
8.1 Representation
8.2 Definition
8.3 Learning from Instructions
8.4 Example: The TICS Architecture
8.5 Limitations
9 Modeling Human Teaching Strategies
9.1 Natural Human Teaching Biases
9.2 A Noisily-Rational Decision Model
10 Applications
10.1 Interactive Task Learning
10.2 Learning Robot Behaviors from Human Demonstrations
10.3 Learning Robot Behaviors from Human Instructions
10.4 Socially Assistive Robotics
11 Conclusions, Challenges and Perspectives
References
Explainable AI
Explainable for Trustworthy AI
1 Introduction
2 Explainable AI
2.1 Types of Explanations
2.2 Desiderata of an Explanation
3 XAI Methods
3.1 Post-hoc Explanations Methods
3.2 Text Data
3.3 Time Series
3.4 Post-hoc Explanations: Hands-on the Code
3.5 Transparent by Design Methods
3.6 Tabular Data
3.7 Image Data
3.8 Text Data
3.9 Interpretable by Design Explanations: Hands-on the Code
4 Conclusion and Open Challenges
References
Why and How Should We Explain AI?
1 Introduction
1.1 Learning Objectives
2 Why Should We Explain Algorithms?
2.1 Robbins: The Impacts of AI and a Catch-22
2.2 London: A Lack of Explanations Elsewhere
2.3 Durán and Jongsma: Reliabilism and Explainability
2.4 Krishnan: Reasons to Explain
3 What Constitutes an Explanation?
3.1 Causal/Interventionist
3.2 Unificationist
3.3 Mechanists
3.4 Bringing Theory into Practice
4 Conclusion
5 Suggested Readings
References
Artificial Visual Intelligence
1 Introduction
2 Application: Human-Centred Cognitive Technologies
3 Deep Semantics: On Neurosymbolic Explainability
3.1 What is Deep Semantics?
3.2 Commonsense, Space, Change
4 Visuospatial Commonsense: On Neurosymbolic Reasoning and Learning
5 Reflection and Outlook
A Select Further Readings
B Visual Computing Foundations
References
Ethics, Law and Society AI
An Introduction to Ethics and AI
1 Introduction
2 Ethics of AI (by Guido Boella)
2.1 The Bias Problem
2.2 Explainability
2.3 Control and Accountability of Autonomous Systems
2.4 Power and the Rhetoric Behind AI
2.5 AI and Economics
3 How AI is Influencing Morality and Ethics (by Maurizio Mori)
4 AI and Ethics (by Guido Boella)
5 Conclusion
Law for Computer Scientists
1 Introduction
1.1 Learning Objectives
1.2 Law and Computing Architectures
2 How Law Works
2.1 What Law is (Not)
2.2 What Law Does (Not)
2.3 Legal Norms
2.4 Legal Rights
2.5 Legal Reasoning
3 Data Protection
3.1 GDPR
3.2 Objectives
3.3 The Object of Regulation
3.4 Addressees: Controller and Processor
3.5 Principles
3.6 Legal Basis
3.7 Data Protection Impact Assessment
3.8 Data Protection by Design and by Default
4 Artificial Intelligence Act
4.1 The Proposal
4.2 Defining AI Systems
4.3 Architecture of the Act
4.4 Qualification of AI Systems
4.5 Final Note on Actionable Protection in the AI Act
References
Mythical Ethical Principles for AI and How to Attain Them
1 About This Tutorial
2 Introduction
3 Ethical Principles
4 Algorithmic Accountability
4.1 What is It?
4.2 Alignment with Ethical Principles
5 Transparency, Explainability, Interpretability
5.1 Transparency
5.2 Explainability and Interpretability
5.3 Information Availability and Ethical Principles
6 Fairness
6.1 Algorithmic Fairness, Bias and Discrimination Definitions
6.2 Fairness and Ethical Principles
7 Privacy
7.1 Definitions and Concerns
7.2 Privacy and Ethics
8 Summary
9 Recommended Further Reading
References
Operationalising AI Ethics: Conducting Socio-technical Assessment
1 Introduction
2 Background
2.1 What Is Responsible AI?
2.2 AI Governance
2.3 Transparency and Control
2.4 The End Game: Accountability
3 Protostrategos
3.1 Scenario Overview
3.2 Students' Reflections
4 Socio-ethical Values in Design and Development
4.1 The Glass-Box Approach
4.2 Interrogating the Black Box
4.3 RAIN
5 Conclusion
6 Further Reading
References
Writing Science Fiction as an Inspiration for AI Research and Ethics Dissemination
1 Learning Objectives
2 Science and Fiction: Mutual Inspiration
3 Fostering the Confluence of Technoscientists and Writers/Artists
4 New Ethics Issues Raised by Artificial Intelligence
5 Introducing Techno-Ethics in the Curricula
6 Science Fiction Narrative Engages Technology Students
7 Question Answering - First Round
8 Plotting, Writing and Teaching with The Vestigial Heart
8.1 Designing the `Perfect' Assistant
8.2 Robot Appearance and Emotion
8.3 Robots in the Workplace
8.4 Robots in Education
8.5 Human-Robot Interaction and Human Dignity
8.6 Social Responsibility and Robot Morality
9 Concluding Remarks and Other Writing Initiatives
9.1 Further Reading
10 Question Answering - Second Round
References
Argumentation
Argumentation in AI
1 Learning Objectives
2 Introduction
3 Abstract Argumentation
3.1 Complexity of Verifying Extensions
4 Uncertainty in Argumentation
4.1 Probabilistic Abstract Argumentation Framework of Form IND
4.2 Probabilistic Abstract Argumentation Framework of Form EX
4.3 Comparison Between IND and EX
4.4 Other Options for Handling Uncertainty
5 Conclusion
References
Computational Argumentation & Cognitive AI
1 Introduction
2 Structured Argumentation
2.1 The GORGIAS Argumentation Framework
3 Cognitive Argumentation
3.1 The Suppression Task
3.2 The COGNICA System
3.3 Argumentation and Cognitive Architectures
4 Argumentation for Learning
4.1 What Should a Language of Learning Be Like?
4.2 Argumentation as a Language of Learning
4.3 Case Study 1: Autodidactic Learning of Arguments
4.4 Case Study 2: eXplanations In, eXplanations Out
5 Applications of Argumentation
5.1 SoDA: Software Development Through Argumentation
5.2 Application Language Levels: Example Applications
5.3 Machine Learning Assisted Policy Formation
References
Social Simulation
Agent-Based Social Simulation for Policy Making
1 Introduction
2 Session 1: Agent-Based Social Simulation
2.1 Foundations of Modeling and Simulation
2.2 Modeling Human Behavior
2.3 ASSOCC - A Simulation Model of the Covid-19 Pandemic
2.4 Practical Exercise: Experimenting with the ASSOCC Model
2.5 Activity Outcome
3 Session 2: Advanced Agent Architectures for Modeling Policies
3.1 Human Behavior as Key for Simulation Policies
3.2 Modeling Human Values, Motivations, and Needs
3.3 Opportunities of Social Simulations
3.4 Practical Exercise: Extending Models to Simulate Policies
3.5 Activity Outcome
4 Session 3: Simulating for Policy Makers
4.1 Conceptual Activity
4.2 Practical Activity: Data to Factors
4.3 Activity Outcome
5 Conclusions
References
Towards a Social Artificial Intelligence
1 Introduction
2 Network Effects of AI and Their Impact on Society
2.1 Emergent Properties of Real Networks
2.2 AI Pitfalls on Real Networks
2.3 Addressing AI Pitfalls
3 Beyond Data-Driven Approaches: Behavioural Models to Understand Societal Phenomena
3.1 Tracking Online Extremism that Leads to Offline Extremist Behaviours
3.2 Multi-agent Behavioural Models for Epidemic Spreading: From Model to Data in the Covid-19 Crisis
4 Hands-on Tutorial
5 Conclusions
References
Author Index
