Transparency and Interpretability for Learned Representations of Artificial Neural Networks

Contents
List of Figures
List of Tables
1 Introduction
	1.1 Object of Investigation
	1.2 Research Questions
	1.3 Structure of the Thesis
2 Background & Foundations
	2.1 A Short History of AI Research
		2.1.1 The Early Years
		2.1.2 The Golden Ages
		2.1.3 The AI Winter
		2.1.4 The AI Renaissance
	2.2 The Modern Era of AI Research
		2.2.1 Deep Learning for Computer Vision
		2.2.2 Computer Vision Beyond the ILSVRC
		2.2.3 From Supervised Learning to Reinforcement Learning
		2.2.4 Deep Reinforcement Learning Breakthroughs for Video and Board Games
		2.2.5 Tackling Games with Imperfect Information
	2.3 Towards Research on Transparency & Interpretability
		2.3.1 A Shift of Paradigm: From Optimizing to Understanding
		2.3.2 Inspirations from Neuroscience Research
3 Methods and Terminology
	3.1 Learned Representations
		3.1.1 Investigating Learned Representations
		3.1.2 Visualizing Structure of the Learned Representations
	3.2 Delimitation of the Object of Investigation
		3.2.1 Transparency for Computer Vision Models
		3.2.2 Transparency for Motion Control Models
		3.2.3 Transfer to an Industrial Application Scenario
4 Related Work
	4.1 Relationship Between a Network’s Input Features and its Output
	4.2 Visualization of Network Properties and Graphical User Interfaces
	4.3 Investigating the Importance of Individual Network Components
		4.3.1 Miscellaneous Contributions
		4.3.2 Ablations Studies
		4.3.3 Reverse Engineering of Neural Networks
5 Research Studies
	5.1 Investigating Learned Representations in Computer Vision
		5.1.1 Research Study 1: Characterizing Single Neurons in a Shallow MLP
		5.1.2 Research Study 2: Network Ablations in a Deep Neural Network
		5.1.3 Research Study 3: Functional Neuron Populations in Custom-made CNNs
	5.2 Investigating Learned Representations in Motor Control
		5.2.1 Research Study 4: Influence of Network Ablations on Activation Patterns
		5.2.2 Research Study 5: Relation Between Neural Activations and Agent Behavior
6 Transfer Studies
	6.1 Transfer Study 1: Network Ablations for Deep Drawing
		6.1.1 Key Contributions of the Study
		6.1.2 Methods and Experimental Design
		6.1.3 Results
		6.1.4 Summary and Contribution of the Results to the Research Questions
	6.2 Transfer Study 2: Attention Mechanisms for Deep Drawing
		6.2.1 Key Contributions of the Study
		6.2.2 Methods and Experimental Design
		6.2.3 Results
		6.2.4 Summary and Contribution of the Results to the Research Questions
7 Critical Reflection & Outlook
	7.1 Reflection of Results & Contribution to Research Questions
		7.1.1 Research Question 1
		7.1.2 Research Question 2
		7.1.3 Research Question 3
		7.1.4 Research Question 4
	7.2 Future Research Directions
8 Summary
References