Artificial Intelligence: What Is Behind the Technology of the Future?

Preface
	Preface to the German Edition
	Preface to the English Edition
	Acknowledgements
Contents
About the Authors
1 What Is Intelligent About Artificial Intelligence?
	1.1 Human Intelligence Has Many Dimensions
	1.2 How To Recognize Artificial Intelligence
	1.3 Computers Learn
	1.4 Deep Neural Networks Can Recognize Objects
	1.5 How To Understand Artificial Intelligence
	1.6 The History of Artificial Intelligence
	1.7 Summary
	References
2 What Are the Capabilities of Artificial Intelligence?
	2.1 Object Recognition in Images
		2.1.1 Medical Diagnosis
		2.1.2 Predicting the 3D Structure of Proteins
	2.2 Speech Recognition
	2.3 Machine Translation
	2.4 Answering Natural Language Questions
	2.5 Dialogs and Personal Assistants
	2.6 Board Games
		2.6.1 The Strategy Game Go
		2.6.2 Artificial Intelligence Wins Against Five Poker Professionals
	2.7 Video Games
		2.7.1 Atari 2600 Game Console
		2.7.2 Capture the Flag Quake
		2.7.3 The Real-Time Strategy Game Dota2
	2.8 Self-Driving Cars
		2.8.1 Further Development of Self-Driving Cars
	2.9 The Computer as a Creative Medium
		2.9.1 Composing New Images
		2.9.2 Inventing Stories
	2.10 General Artificial Intelligence
	2.11 Summary
	References
3 Some Basic Concepts of Machine Learning
	3.1 Main Types of Machine Learning
		3.1.1 Supervised Learning
		3.1.2 Unsupervised Learning
		3.1.3 Reinforcement Learning
	3.2 Programming and Learning
		3.2.1 Models Transform an Input into an Output
		3.2.2 Algorithms Process a List of Instructions Step by Step
		3.2.3 A Learning Problem: The Recognition of Digits
		3.2.4 Vectors, Matrices and Tensors
	3.3 Learning a Relationship
		3.3.1 Scheme for Learning: Model, Loss Function and Optimization
		3.3.2 Detailed Process of Learning
	3.4 A Simple Model: Logistic Regression
		3.4.1 Calculation of a Score
		3.4.2 The Simultaneous Calculation of All Scores
		3.4.3 Affine Transformation
		3.4.4 The Softmax Function Generates a Probability Vector
		3.4.5 The Logistic Regression Model
	3.5 Measuring Model Performance
		3.5.1 A Criterion of Model Performance: The Likelihood of Complete Training Data
		3.5.2 How to Measure Learning Success: The Loss Function
		3.5.3 Illustration for Two Classes and Two Input Features
	3.6 Optimization, or How to Find the Best Parameter Values
		3.6.1 The Gradient Indicates the Direction of the Steepest Ascent
		3.6.2 The Gradient for Several Dimensions
		3.6.3 The Gradient of the Loss Function
		3.6.4 Stepwise Minimization by Gradient Descent
		3.6.5 The Learning Rate Sets the Length of an Optimization Step
		3.6.6 Minibatch Gradient Descent Needs Less Computation
		3.6.7 Applying the Model to New Data
		3.6.8 Checking the Accuracy on the Test Set
		3.6.9 Precision and Recall for Classes of Different Size
	3.7 Summary
	References
4 Deep Learning Can Recognize Complex Relationships
	4.1 The XOR Problem Involves Interactions Between Features
	4.2 Nonlinearities Create Curved Separating Planes
	4.3 Deep Neural Networks Are Stacks of Nonlinear Layers
		4.3.1 Vectors and Tensors Represent the Transformed Contents
	4.4 Training of DNN with the Backpropagation Approach
	4.5 Toolkits Facilitate DNN Specification and Training
		4.5.1 Parallel Computations Accelerate DNN Training
		4.5.2 Toolkits Simplify Work with DNN
	4.6 How to Improve the Network?
		4.6.1 Iterative Model Construction Using the Validation Set
		4.6.2 Underfitting and Overfitting Lead to Higher Errors
		4.6.3 An Example of Overfitting
		4.6.4 Regularization Procedures Reduce Overfitting Errors
		4.6.5 Penalizing Large Parameter Values Reduces Abrupt Output Changes
		4.6.6 Dropout Disables Parts of the Network
		4.6.7 Batch Normalization Avoids Extreme Values of Hidden Vectors
		4.6.8 Mathematical Proof: Stochastic Gradient Descent Finds Well Generalizing DNN
	4.7 Different Applications Require Different Networks Structures
		4.7.1 Multilayer Feedforward Network
		4.7.2 Convolutional Neural Network (CNN)
		4.7.3 Recurrent Neural Network (RNN)
		4.7.4 Reinforcement Learning Network
		4.7.5 Generative Adversarial Network (GAN)
		4.7.6 Autoencoder Networks Produce a Compressed Representation
		4.7.7 Architectures for Specific Media and Application Areas
	4.8 The Design of a Deep Neural Network Is a Search Process
		4.8.1 Selection of the Hyperparameters of the Network
		4.8.2 The Standard Model Search Process Leads to Better Models
		4.8.3 Automatic Search of Model Architectures and Hyperparameters
	4.9 Biological Neural Networks Work Differently
	4.10 Summary and Trends
	References
5 Image Recognition with Deep Neural Networks
	5.1 What Does Image Recognition Actually Mean?
		5.1.1 Types of Object Recognition in Images
		5.1.2 Inspirations from Biology
		5.1.3 Why Is Image Recognition Difficult?
	5.2 The Components of a Convolutional Neural Network
		5.2.1 Convolutional Kernels Analyze Small Image Areas
		5.2.2 Different Kernels in a Convolution Layer Compute Many Features
		5.2.3 The Pooling Layer Selects Most Important Feature Values
	5.3 A Simple Convolutional Neural Network for Digit Recognition
	5.4 ImageNet Competition Boosts Method Development
	5.5 Advanced Convolutional Neural Networks
		5.5.1 AlexNet Successfully Uses GPUs for Training
		5.5.2 ResNet Facilitates Optimization by Residual Connections
			ResNet Requires Enormous Computing Power
		5.5.3 DenseNet Employs Additional Residual Connections
		5.5.4 Transformed Images Improve ResNeXt Training
	5.6 Analysis of CNN Results
		5.6.1 Individual Kernels Respond To Features of Different Types and Sizes
		5.6.2 Similar Images Correspond To Neighboring Hidden Vectors
	5.7 Transfer Learning Reduces the Need for Training Data
	5.8 Localization of Objects in an Image
		5.8.1 Object Localization by Rectangles
		5.8.2 Pixel-Precise Localization of Class Objects
		5.8.3 Max-Unpooling Assigns Values To an Enlarged Field
		5.8.4 U-Net Detects Objects and Then Finds the Associated Pixels
	5.9 3D Reconstruction of a Scene
	5.10 Human Faces Can Be Matched with High Accuracy
	5.11 Assessing the Accuracy of Model Predictions
		5.11.1 Uncertainty of Model Predictions
		5.11.2 Bootstrap Generates a Set of Plausible Models
		5.11.3 Bayesian Neural Networks
	5.12 Reliability of Image Recognition
		5.12.1 Influence of Image Distortions
		5.12.2 Targeted Construction of Misclassified Images
	5.13 Summary and Trends
	References
6 Capturing the Meaning of Written Text
	6.1 How to Represent the Meaning of Words by Vectors?
		6.1.1 The Concept of Embedding Vectors
		6.1.2 Computation of Embedding Vectors with Word2vec
		6.1.3 Softmax Function Approximation Reduces Computation
	6.2 Properties of Embeddings Vectors
		6.2.1 Nearest Neighbors of Embeddings Have Similar Meanings
		6.2.2 Differences Between Embeddings Express Relations
		6.2.3 FastText Uses N-Grams of Letters
		6.2.4 StarSpace Creates Embeddings for Other Objects
	6.3 Recurrent Neural Networks for Sequence Modeling
		6.3.1 Recurrent Neural Networks as Language Models
		6.3.2 Training of Recurrent Neural Networks
		6.3.3 The Properties of RNN Gradients
	6.4 The Long Short-Term Memory (LSTM) is a Long-Term Memory
		6.4.1 Controling Memory Operations by Gates
		6.4.2 LSTMs with Multiple Layers
		6.4.3 Applications of the LSTM
		6.4.4 Bidirectional LSTM Networks for Word Property Prediction
		6.4.5 Visualization of Recurrent Neural Networks
	6.5 Transformation of one Sequence into Another Sequence
		6.5.1 Sequence-to-Sequence Networks for Translation
			Creating and Evaluating a Translation
		6.5.2 Attention: Improving Translation by Recourse to Input Words
		6.5.3 Attention Generated Translation Results
	6.6 BERT: A Model for the Representation of Meanings
		6.6.1 Tokenization to Limit Vocabulary Size
		6.6.2 Self-Attention Analyzing the ``Correlation\'\' of Different Tokens
		6.6.3 BERT Computes Contextual Embeddings by Self-Attention
		6.6.4 BERT Prediction Tasks for Unsupervised Pre-training
	6.7 Transfer Learning with BERT Language Models
		6.7.1 Semantic Classification Tasks
		6.7.2 Question Answering
		6.7.3 Extraction of World Knowledge
		6.7.4 Using BERT for Web Search
	6.8 Transformer Translation Models
		6.8.1 Cross-Attention Exploits the Input-Output ``Correlation\'\'
		6.8.2 Transformer Architecture Uses Self- and Cross-Attention
		6.8.3 Training the Transformer for Language Translation
		6.8.4 Translation Results for the Transformer Model
		6.8.5 Simultaneous Translation Requires a Time Delay
		6.8.6 Transfer Learning for Translation Models
	6.9 The Description of Images by Text
		6.9.1 Explanation of DNN Forecasts
		6.9.2 Explanations Are Necessary
		6.9.3 Global Explanatory Models
		6.9.4 Local Explanatory Models
	6.10 Reliability of Text Understanding
		6.10.1 Robustness in Case of Text Errors and Domain Change
		6.10.2 Vulnerability to Malicious Modification of Inputs
	6.11 Summary and Trends
	References
7 Understanding Spoken Language
	7.1 Speech Recognition
		7.1.1 Why Is Speech Recognition Difficult?
		7.1.2 How to Represent Speech Signals in the Computer?
		7.1.3 Assessing Speech Recognition Accuracy
			The Word Error Rate WER
			Established Speech Recognition Benchmarks
		7.1.4 The History of Speech Recognition
	7.2 Deep Sequence-to-Sequence Models
		7.2.1 List-Attend-Spell Generates a Sequence of Letters
		7.2.2 Sequence-to-Sequence Model for Wordsand Syllables
	7.3 Convolutional Neural Networks for Speech Recognition
		7.3.1 CNN Models
		7.3.2 Combined Models
			ResNet and BiLSTM
			Augmentation of Training Data
	7.4 Lip Reading
	7.5 Generating Spoken Language from Text
		7.5.1 WaveNet with Dilated Convolution for Long Dependencies
		7.5.2 The Tacotron Generates a Spectrogram
	7.6 Dialogs and Voice Assistants
	7.7 Gunrock: An Extended Alexa Voice Assistant
		7.7.1 Language Understanding
		7.7.2 Dialog Management
		7.7.3 Response Generation
		7.7.4 Testing the Voice Assistant
	7.8 Analysis of the Content of Videos
		7.8.1 Tasks of Video Content Analysis
		7.8.2 Training Data for Classification of Videosby Activities
		7.8.3 Convolution Layers for Video Content Recognition
		7.8.4 Accuracy of Video Classification
		7.8.5 The Generation of Subtitles for Videos
	7.9 Reliability of Processing Spoken Language
		7.9.1 The Effect of Noise and Other Distortions on Speech Recognition
		7.9.2 Adversarial Attacks on Automatic SpeechRecognition
	7.10 Summary
	References
8 Learning Optimal Policies
	8.1 Some Basic Definitions
	8.2 Deep Q-Network
		8.2.1 Policy to Maximize the Sum of Rewards
		8.2.2 A Small Navigation Task
		8.2.3 Discounted Future Reward Encourages Fast Solutions
		8.2.4 The Q-Function Evaluates State-Action Pairs
		8.2.5 The Bellman Equation Relates Q-Valuesto Each Other
		8.2.6 Approximation of the Q-Function by a Deep Neural Network
		8.2.7 Q-Learning: Training a Deep Q-Network
			Creating an Episode with the Deep Q-Network
			Optimization with the Generated Episode
			Practical Tricks: Selection of Training Examples and Loss Function Calculation
			Exploration
	8.3 Application of Q-Learning to Atari Video Games
		8.3.1 Definition of the Game State in Atari Games
		8.3.2 Architecture of the Atari Q-Network
		8.3.3 Training of Atari Q-Networks
		8.3.4 Evaluation of Atari Q-Networks
	8.4 Policy Gradients for Learning Stochastic Policies
		8.4.1 Need for Policies with Random Elements
		8.4.2 Direct Optimization of a Policy by Policy Gradients
		8.4.3 Extensions of the Policy Gradient: Actor-Critic and Proximal Policy Optimization
		8.4.4 Application to Robotics and the Go Board Game
		8.4.5 Application to the Dota2 Real-Time Strategy Game
	8.5 Self-Driving Cars
		8.5.1 Sensors of Self-Driving Cars
		8.5.2 Functionality of an Agent for Autonomous Driving
		8.5.3 Fine-Tuning Through Simulation
		8.5.4 Reliability of Reinforcement Learning
		8.5.5 Simulation-Trained Models Difficult to Transfer
		8.5.6 Adversarial Attacks on Reinforcement LearningModels
	8.6 Summary and Trends
	References
9 Creative Artificial Intelligence and Emotions
	9.1 Image Creation with Generative Adversarial Networks (GAN)
		9.1.1 Forger and Art Expert
		9.1.2 Generator and Discriminator
		9.1.3 Optimization Criteria for Generator andDiscriminator
		9.1.4 Results of Generative Adversarial Networks
		9.1.5 Interpolation Between Images
		9.1.6 Transformation of Images
		9.1.7 Transformation of Images Without Training Pairs
		9.1.8 Creative Adversarial Network
		9.1.9 Generating Images from Text
		9.1.10 GAN-Generated Persons in Three Dimensions
	9.2 Composing Texts
		9.2.1 Automatic Reporter: Convert Data to Newspaper Reports
		9.2.2 Generating Longer Stories
		9.2.3 GPT-2 Invents Complex Stories
			Generative Capability of GPT-2
			Visualization of Model Predictions
			Further Developments of GPT-2
	9.3 Compose Music Automatically
		9.3.1 MuseNet Composes Mixtures of Classic and Pop
		9.3.2 The Music Transformer Invents Piano Pieces
	9.4 Emotions and Personality
		9.4.1 A XiaoIce Dialog
		9.4.2 The Goal: Encourage People to Keep Talking
		9.4.3 Architecture of XiaoIce
		9.4.4 Number of User Responses as Optimization Criterion
		9.4.5 Emotional Empathy and Support
		9.4.6 Summary and Trends
	References
10 AI and Its Opportunities, Challenges and Risks
	10.1 Opportunities for Economy and Society
		10.1.1 Smart Home, My House Takes Care of Me
			What Is the Benefit for the User?
		10.1.2 Diagnosis, Therapy, Care and Administration in Medicine
			AI in Early Detection and Diagnosis
			AI in Therapy
			AI in Nursing
			AI in Hospital Administration
		10.1.3 Machine Learning in Industrial Applications
		10.1.4 Further Areas of Application for AI
	10.2 Economic Impacts and Interrelationships
		10.2.1 The Monetization of Data
		10.2.2 The New Digital Service World: AI as a Service
		10.2.3 Large Companies as Drivers of AI
			The AI Company Google
			The Chinese Competitors
			Chance for Europe
		10.2.4 Impact on the Labor Market
			Education for a Digital World
			The Job Description of the Data Scientist
	10.3 Challenges for the Society
		10.3.1 Challenges of AI in Medicine
		10.3.2 Orwell\'s 1984 Vers. 2.0: AI as a Surveillance Tool
		10.3.3 War of the Machines
		10.3.4 Artificial General Intelligence
	10.4 Methodological Challenges
		10.4.1 Combination of Data and Uncertain Reasoning
		10.4.2 Fast and Slow Thinking
	10.5 Building Trust in AI
		10.5.1 How to Build Trustworthy AI Systems?
		10.5.2 How to Test Deep Neural Networks?
		10.5.3 Is Self-Determined, Effective Use of an AI System Possible?
		10.5.4 Does the AI System Treat all Affected Parties Fairly?
		10.5.5 Are the Functioning and Decisions of AI Comprehensible?
		10.5.6 Are AI Systems Secure from Attack, Accident, and Error?
		10.5.7 Do AI Components Work Reliably and Perform Robustly?
		10.5.8 Does AI Protect Privacy and Other Sensitive Information?
		10.5.9 The Challenges for an AI Seal of Approval
	10.6 Summary
	References
A Appendix
	A.1 Mathematical Notation
	A.2 Glossary
	A.3 List of Images and Their Sources
References
Index