Interpretability in Deep Learning

1. Auflage 2023
ISBN: 978-3-031-20638-2
Verlag: Springer

Buch, Englisch, 466 Seiten, Format (B × H): 160 mm x 241 mm, Gewicht: 981 g

This book is a comprehensive curation, exposition and illustrative discussion of recent research tools for interpretability of deep learning models, with a focus on neural network architectures. In addition, it includes several case studies from application-oriented articles in the fields of computer vision, optics and machine learning related topic. 

The book can be used as a monograph on interpretability in deep learning covering the most recent topics as well as a textbook for graduate students. Scientists with research, development and application responsibilities benefit from its systematic exposition.

Zielgruppe

Research

Autoren/Hrsg.

Somani, Ayush

Horsch, Alexander

Prasad, Dilip K.

Fachgebiete

• Wirtschaftswissenschaften Betriebswirtschaft Unternehmensforschung
• Wirtschaftswissenschaften Betriebswirtschaft Management Wissensmanagement
• Mathematik | Informatik EDV | Informatik Informatik Bildsignalverarbeitung
• Mathematik | Informatik EDV | Informatik Informatik Künstliche Intelligenz Computer Vision

Inhaltsverzeichnis

1 INTRODUCTION

1.1 Deep Learning Glossary

1.2 Evolution of Deep Learning

1.2.1 Neural Learning

1.2.2 Fuzzy Learning

1.2.3 Convergence of Fuzzy Logic and Neural Learning

1.2.4 Synergy of Neuroscience and Deep Learning

1.3 Awakening of Interpretability

1.3.1 Relevance

1.3.2 Necessity

1.3.3 The Taxonomy of Interpretability

1.4 The Question of Interpretability

1.4.1 Interpretability - Metaverse

1.4.2 Interpretability - The Right Tool

1.4.3 Interpretability - The Wrong Tool

2 NEURAL NETWORKS FOR DEEP LEARNING

2.1 Neural Network Architectures

2.1.1 Perceptron

2.1.2 Artificial Neural Network

2.1.3 Recurrent Neural Network

2.1.4 Convolutional Neural Network

2.1.5 Autoencoder Neural Network

2.1.6 Generative Adversarial Network

2.1.7 Graph Neural Network

2.2 Learning Mechanisms

2.2.1 Activation function

2.2.2 Forward Propagation

2.2.3 Backpropagation

2.2.4 Gradient Descent

2.2.5 Learning Rate

2.2.6 Optimization

2.2.7 Initialization

2.2.8 Regularization

2.3 Challenges and Limitations of Traditional Techniques

2.3.1 Resource-Demanding Checks

2.3.2 Uncertainty Measure

2.3.3 Network Learning Sanity Check

2.3.4 Gradient Checks

2.3.5 Decision Transparency

3 KNOWLEDGE ENCODING AND INTERPRETATION

3.1 What is Knowledge?

3.1.1 Image Representation

3.1.2 Word Representation

3.1.3 Graph Representation

3.2 Knowledge Encoding and Architectural Understanding

3.2.1 Role of Neurons

3.2.2 Role of Layers

3.2.3 Role of Explanation

3.2.4 Semantic Understanding

3.2.5 Network Understanding

3.3 Design and Analysis of Interpretability

3.3.1 Divide and Conquer

3.3.2 Greedy

3.3.3 Back-tracking

3.3.4 Dynamic

3.3.5 Branch and Bound

3.3.6 Brute-force

3.4 Knowledge Propagation in Deep Network Optimizers

3.4.1 Knowledge versus Performance

3.4.2 Deep versus Shallow Encoding

4 INTERPRETATION IN SPECIFIC DEEP ARCHITECTURES

4.1 Interpretation in Convolution Networks

4.1.1 Case Study: Image Representation by Unmasking Clever Hans

4.1.2 Variants of CNNs

4.1.3 Interpretation of CNNs

4.1.4 Review: CNN Visualization Techniques

4.1.5 Review: CNN Adversarial Techniques

4.1.6 Inverse Image Representation

4.1.7 Case study: Superpixels Algorithm

4.1.8 Activation Grid and Activation Map

4.1.9 Convolution Trace

4.2 Interpretation in Autoencoder Networks

4.2.1 Visualization of Latent Space

4.2.2 Sparsity and Interpretation

4.2.3 Case Study: Microscopy Structure to Structure Learning

4.3 Interpretation in Adversarial Networks

4.3.1 Interpretation in Generative Network

4.3.2 Interpretation in Latent Spaces

4.3.3 Evaluation Metrics

4.3.4 Case study: Digital Staining of Microscopy Images

4.4 Interpretation in Graph Networks

4.4.1 Neural Structured Learning

4.4.2 Graph Embedding and Interpretability

4.4.3 Evaluation Metrics for Interpretation

4.4.4 Disentangled Representation Learning on Graphs

4.4.5 Future Direction

4.5 Self-Interpretable Models

4.6 Pitfalls of Interpretability Methods

5 FUZZY DEEP LEARNING

5.1 Fuzzy Theory

5.1.1 Fuzzy Sets and Fuzzy Membership

5.1.2 Fuzzification and Defuzzification

5.1.3 Fuzzy Rules and Inference Systems

5.2 Neuro-Fuzzy Inference Systems

5.2.1 Architecture of a Neuro-Fuzzy Inference System

5.2.2 Other Design Elements of Neuro-Fuzzy Inference Systems

5.2.3 Learning mechanisms for Neuro-Fuzzy Inference Systems

5.2.4 Online Learning with Dynamic Streaming Data

5.3 Case studies

5.3.1 POPFNN Family of NFS - evolution towards sophisticated brain-like

learning

5.3.2 Combining Conventional Deep Learning and Fuzzy Learning

A Mathematical models and theories

A.1 Choquet Integral

A.1.1 Restricting the Scope of FM/ChI

A.1.2 ChI Understanding from NN

A.2 Deformation Invariance Property

A.3 Distance Metrics

A.4 Grad Weighted Class Activation Mapping

A.5 Guided Saliency

A.6 Jensen-Shanon Divergence

A.7 Kullback-Leibler Divergence

A.8 Projected Gradient Descent

A.9 Pythagorean Fuzzy Number

A.10 Targeted Adversarial Attack

A.11 Translation Invariance Property

A.12 Universal Approximation Theorem

A List of digital resources and examples

References .

Über Autor(innen)

Ayush Somani is a research fellow in the Department of Computer Science at UiT The Arctic University of Norway. He received his Integrated Masters from Indian Institute of Technology (ISM) Dhanbad in Maths and Computing. He has earned multiple honors like Dare2Compete Awards 2021, Samsung Innovation Award 2020, KDD'20 Travel Award, and Finalist in Machine Learning & Software Development Flipkart Grid 2.0 challenge. At Travel Buddy, he worked as a data scientist intern to implement AI-automated content moderation. He has research interests in interpretability, explanability, and other aspects of deep learning.

Alexander Horsch, born 1955, is a full professor at the Department of Computer Science, UiT The Arctic University of Norway. He holds a Ph.D. in Computer Science (1989) and a Dr. med. habil in Medical Informatics (1999), both from the Technical University of Munich (TUM). He was the head of the Medical Computing Center at Klinikum rechts der Isar, TUM (1987–1995), and researcher and lecturer, later associate professor and APL professor at TUM Medical Faculty (1996–2015). Several research projects in telemedicine and computer-aided diagnosis with grants from the German Ministry of Research and Technology, the Bavarian State Government, the European Union, and the German Telekom have been managed by him. From beginning 2015 to summer 2019, he was the head of the Department of Computer Science at UiT. He is a member of the research group for physical activity at UiT Medical Faculty, focusing on sensor data analysis within the Tromsø Study, a large epidemiological trial. He is the principal investigator of the interdisciplinary project VirtualStain (2020–2024) at UiT. Earlier, he has worked with the European Space Agency (ESA) since 2004 when he was a member of the ESA Telemed Working Group and with the World Health Organization (WHO) since 2005 as an eHealth and telemedicine expert. From 2006, he has worked in different periods as a consultant for EC (Telemedicine Task Force) and ESA in the Satellite-Enhanced eHealth for sub-Saharan Africa (eHSA) program. Since 2011, he was also supporting the United Nations Office for Outer Space Affairs (UNOOSA) in its Human Space Technology Initiative (HSTI). He is author or co-author of numerous scientific publications and has supervised a dozen doctoral students. His professional expertise ranges from eHealth applications to medical decision support. He has led or was a partner in projects for teleservices in gastroenterology and other medical specialties, web-based multi-modal interactive teaching of tumor diagnostics, case-based ophthalmologic eLearning, early detection of malignant melanoma, quantitative measurement of tumors using tomography data, and accelerometry for physical activity measurements in population studies and clinical research. His current scientific focus is on data analytics applied to biosensor time series and biological images using classical and machine learning approaches.

Dilip K. Prasad is an associate professor in the Department of Computer Science at UiT The Arctic University of Norway. He received the Ph.D. from Nanyang Technological University, Singapore and B.Tech. degree in Computer Science and Engineering from Indian Institute of Technology (ISM) Dhanbad, India. He was a senior research fellow at Nanyang Technological University, Singapore and research fellow at National University of Singapore. He has 5years of industrial experience with IBM, Infosys, Mediatek and Philips. He was a Kauffman Global Scholarship fellow in 2011. He has received 'Rolls-Royce Inventor Award' and several research grants from European Union, Research Council Norway and UiT The Arctic University of Norway. He is a founding member of Bio-AI Research Group at UiT The Arctic University of Norway.  His research interests include image processing, pattern recognition, computer vision and artificial intelligence. He is passionate about making artificial intelligence interpretable and scalable toward bridging the intelligence gap between human and machines.