2019 ICCV Workshop on

Interpreting and Explaining Visual Artificial Intelligence Models


Saturday, November 2nd, 2019

@ COEX 308BC, Seoul, Korea

Invitation

Explainable and interpretable machine learning models and algorithms are important topics which have received growing attention from research, application and administration. Many advanced visual artificial intelligence systems are often perceived as black-boxes. Researchers would like to be able to interpret what the AI model has learned in order to identify biases and failure models and improve models. Many government agencies pay special attention to the topic. In May 2018, EU enacted the General Data Protection Regulation (GDPR) which mandates a right to explanation from machine learning model. An USA agency, DARPA, launched an Explainable Artificial Intelligence program in 2017. The ministry of science and ICT (MSIT) of South Korea established an Explainable Artificial Intelligence Center.

Recently, several models and algorithms are introduced to explain or interpret the decision of complex artificial intelligence system. Explainable Artificial Intelligence systems can now explain the decisions of autonomous system such as self-driving cars and game agents trained by deep reinforcement learning. Saliency maps built by attribution methods such as network gradients, DeConvNet, Layer-wise Relevance Propagation, PatternAttribution and RISE can identify relevant inputs upon the decisions of classification or regression tasks. Bayesian model composition methods can learn automated decomposition of input data into a composition of explainable base models in human pose estimation. Model agnostic methods such as LIME (Local interpretable model-agnostic explanations) and SHAP make complex deep learning models by providing importance of input features. Network Dissection and GAN Dissection provide human-friendly interpretations of internal nodes in deep neural network and deep generative models.

The present workshop aims to overview recent advances in explainable/interpretable artificial intelligence and establish new theoretical foundations of interpreting and understanding visual artificial intelligence models including deep neural networks. We will discuss the future research directions and applications of explainable visual artificial intelligence.

This workshop has interest including but not limited to the following topics.
- Explaining the decision of visual deep learning models
- Interpretable deep learning models
- Machine learning/deep learning models which generates human-friendly explanations
- Bayesian model composition/decomposition methods
- Model-agnostic machine learning explainable models
- Evaluation of explainable AI models
- Causal analysis of complex AI/ML systems
- Practical applications of explainable AI

Invited Speakers


Trevor Darrell
Professor
UC Berkeley

Wojciech Samek
Head of Machine Learning Group
Fraunhofer Heinrich Hertz Institute

David Bau
PhD student
MIT

Ludwig Schubert
Software Engineer
OpenAI

Program

PROGRAM

Time Title Room
9:00 - 9:10 Opening Remarks
308BC
9:10 -10:00 Invited Talk 1. "Recent progress towards XAI at UC Berkeley"
      Trevor Darrell (UC Berkeley)
10:00 - 10:20 Oral Presentations 1 (10min * 2papers)
10:20 - 10:40 Coffee Break
10:40 - 11:30 Invited Talk 2. "Meta-Explanations, Interpretable Clustering & Other Recent Developments"
       Wojciech Samek (Fraunhofer Heinrich Hertz Institute)
308BC
11:30 - 11:55 Poster Spotlights (1min * 25papers)
12:00 - 12:50 Lunch
12:50 - 13:20 Poster Session 1 (16papers) E25~40
13:30 - 14:20 Invited Talk 3. "The Role of Individual Units in Deep Networks in Vision"
      David Bau (MIT)
308BC
14:20 -15:00 Oral Presentations 2 (10min * 4papers)
15:00 - 15:30 Poster Session 2 (15papers) E25~40
15:30 - 15:50 Coffee Break
15:50 - 16:40 Invited Talk 4. "Zooming in: From Activation Atlases down to Features & Circuits"
       OpenAI Clarity Team, Ludwig Schubert
308BC
16:40 - 16:50 Tutorial 1. "Tutorial on TorchRay: a PyTorch interpretability library for reproducible research"
      Ruth Fong (University of Oxford)
16:50 - 16:55 Tutorial 2. "An Open Source Repository of Explainable Artificial Intelligence Projects"
      Sohee Cho (Explainable Artificial Intelligence Center)
16:55 - 17:00 Closing Remarks




Presentaion List

Oral Presentation 1 (10:00-10:20)
Paper ID: 18 - Characterizing Sources of Uncertainty to Proxy Calibration and Disambiguate Annotator and Data Bias
Asma Ghandeharioun (MIT)*, Brian Eoff (Google Research), Brendan Jou (Google Research), Rosalind Picard (MIT Media Lab)

Paper ID: 34 - Decision explanation and feature importance for invertible networks
Juntang Zhuang (Yale University)*, Nicha Dvornek (Yale University), Xiaoxiao Li (Yale University), Junlin Yang (Yale University), James S Duncan (Yale University)

Oral Presentation 2 (14:20-15:00)
Paper ID: 19 - Adaptive Activation Thresholding: Dynamic Routing Type Behavior for Interpretability in Convolutional Neural Networks
Yiyou Sun (University of Wisconsin Madison)*, Sathya Ravi (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison USA)

Paper ID: 35 - Free-Lunch Saliency via Attention in Atari Agents
Dmitry Nikulin (Samsung AI Center Moscow)*, Anastasia Ianina (Samsung AI Center Moscow), Vladimir A Aliev (Samsung AI Center, Moscow), Sergey I Nikolenko (PDMI RAS)

Paper ID: 25 - Interpretable BoW Networks for Adversarial Example Detection
Krishna Kanth Nakka (EPFL)*, Mathieu Salzmann (EPFL)

Paper ID: 17 - Leveraging Model Interpretability and Stability to increase Model Robustness
Fei WU (CentraleSupelec), Thomas MICHEL (Valeo)*, Alexandre Briot (Valeo)

Poster Session 1 (13:00-13:30)
Paper ID: 1 - Visualization of Time Series Deep Neural Network
Sohee Cho (UNIST)*, Jaesik Choi (KAIST)

Paper ID: 2 - Why are Saliency Maps Noisy? Cause of and Solution to Noisy Saliency Maps
Beomsu Kim (Korea Advanced Institute of Science and Technology)*, Junghoon Seo (Satrec Initiative), SeungHyun Jeon (Satrec Initiative), Jamyoung Koo (SI Analytics), Jeongyeol Choe (SI Analytics), Taegyun Jeon (SI Analytics)

Paper ID: 6 - Interpretable Disentanglement of Neural Networks by Extracting Class-Specific Subnetwork
Yulong Wang (Tsinghua University), Xiaolin Hu (Tsinghua University), Hang Su (Tsinghua Univiersity)*

Paper ID: 7 - Occlusions for Effective Data Augmentation in Image Classification
Ruth C Fong (University of Oxford)*, Andrea Vedaldi (Oxford University)

Paper ID: 8 - Grid Saliency for Context Explanations of Semantic Segmentation
Lukas Hoyer (Bosch Center for Artificial Intelligence)*, Mauricio Munoz (Bosch Center for Artificial Intelligence), Prateek Katiyar (Bosch Center for Artificial Intelligence), Anna Khoreva (Bosch Center for Artificial Intelligence), Volker Fischer (Bosch Center for Artificial Intelligence)

Paper ID: 9 - Explaining Visual Models by Causal Attribution
Álvaro Parafita Martínez (Universitat de Barcelona)*, Jordi Vitria (Universitat de Barcelona)

Paper ID: 10 - Explaining Convolutional Neural Networks using Softmax Gradient Layer-wise Relevance Propagation
Brian K Iwana (Kyushu University)*, Ryohei Kuroki (Kyushu University), Seiichi Uchida (Kyushu University)

Paper ID: 11 - Understanding Convolutional Networks Using Linear Interpreters (Extended Abstract)
Pablo Navarrete Michelini (BOE Technology Group Co., Ltd.)*, Hanwen Liu (BOE Technology Group Co., Ltd.), Yunhua Lu (BOE Technology Group Co., Ltd.), Xingqun Jiang (BOE Technology Group Co., Ltd.)

Paper ID: 12 - Bin-wise Temperature Scaling (BTS):Improvement in Confidence Calibration Performance through Simple Scaling Techniques
Younghak Shin (LGCNS)*, ByeongmoonJi (LG CNS), Hyemin Jung (LG CNS), jihyeun yoon (LG CNS), kyungyul kim (LG CNS)

Paper ID: 13 - Towards Analyzing Semantic Robustness of Deep Neural Networks
Abdullah J Hamdi (KAUST)*, Bernard Ghanem (KAUST)

Paper ID: 14 - Towards A Rigorous Evaluation Of XAI Methods On Time Series
Udo M Schlegel (University Konstanz)*, Hiba Arnout (Siemens CT & TU Munich), Mennatallah El-Assady (University of Konstanz), Daniela Oelke (Siemens CT), Daniel Keim (Uni. Konstanz)

Paper ID: 15 - Class Feature Pyramids for Video Explanation
Alexandros Stergiou (Utrecht University)*, George Kapidis (Utrecht University), Grigorios Kalliatakis (University of Essex, UK), Christos Chrysoulas (London South Bank University), Ronald Poppe (Utrecht University), Remco C.Veltkamp (Utrecht University)

Paper ID: 16 - Localizing Occluders with Compositional Convolutional Networks
Adam Kortylewski (Johns Hopkins University)*, Qing Liu (Johns Hopkins University), Huiyu Wang (Johns Hopkins University), Zhishuai Zhang (Johns Hopkins University), Alan Yuille (Johns Hopkins University)

Paper ID: 18 - Characterizing Sources of Uncertainty to Proxy Calibration and Disambiguate Annotator and Data Bias
Asma Ghandeharioun (MIT)*, Brian Eoff (Google Research), Brendan Jou (Google Research), Rosalind Picard (MIT Media Lab)

Paper ID: 20 - Efficient Exploration-based Sampling in the Generative Boundary of Deep Generative Neural Networks
Giyoung Jeon (Ulsan National Institute of Science and Technology), Haedong Jeong (Ulsan National Institute of Science and Technology), Jaesik Choi (KAIST)*

Paper ID: 34 - Decision explanation and feature importance for invertible networks
Juntang Zhuang (Yale University)*, Nicha Dvornek (Yale University), Xiaoxiao Li (Yale University), Junlin Yang (Yale University), James S Duncan (Yale University)

Poster Session 2 (15:00-15:30)
Paper ID: 17 - Leveraging Model Interpretability and Stability to increase Model Robustness
Fei WU (CentraleSupelec), Thomas MICHEL (Valeo)*, Alexandre Briot (Valeo)

Paper ID: 19 - Adaptive Activation Thresholding: Dynamic Routing Type Behavior for Interpretability in Convolutional Neural Networks
Yiyou Sun (University of Wisconsin Madison)*, Sathya Ravi (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison USA)

Paper ID: 21 - Cost-Effective Interactive Attention Learning for Action Recognition
Jay Heo (KAIST)*, Junhyeon Park (KAIST), Hyewon Jeong (KAIST), Wuhyun Shin (KAIST), Kwang Joon Kim (Yonsei University College of Medicine), Sung Ju Hwang (KAIST)

Paper ID: 22 - A Plug-in Factorizer for Disentangling a Latent Representation
Jee Seok Yoon (Korea University), Wonjun Ko (Korea University), Heung-Il Suk (Korea University)*

Paper ID: 24 - Visual Understanding of Multiple Attributes Learning Model of X-Ray Scattering Images
Xinyi Huang (Kent State University), SuphanutJamonnak (Kent State University), Ye Zhao (Kent State University)*, Boyu Wang (Stony Brook University), Minh Hoai Nguyen (Stony Brook University), KevinYager (Brookhaven National Laboratory), Wei Xu (Brookhaven National Lab)

Paper ID: 25 - Interpretable BoW Networks for Adversarial Example Detection
Krishna Kanth Nakka (EPFL)*, Mathieu Salzmann (EPFL)

Paper ID: 28 - Semantically Interpretable Activation Maps: what-where-how explanations within CNNs
Diego Marcos (Wageningen University)*, Sylvain Lobry (Wageningen University and Research), Devis Tuia (Wageningen University and Research)

Paper ID: 30 - To Trust, or Not to Trust? A Case Study of Human Bias in Automated Video Interview Assessments
Chee Wee Leong (Educational Testing Service (ETS))*, Katrina Roohr (Educational Testing Service), Vikram Ramanarayanan (University of California, San Francisco), MichelleMartin-Raugh (Educational Testing Service (ETS)), Harrison Kell (Educational Testing Service), Rutuja Ubale (Educational Testing Service Research), Yao Qian (Educational Testing Service), Zydrune Mladineo (Educational Testing Service), Laura McCulla (Educational Testing Service)

Paper ID: 31 - Assisting human experts in the interpretation of their visual process: A case study on assessing copper surface adhesive potency
Tristan Hascoet*, Xuejiao Deng, Kiyoto Tai, Mari Sugiyama, Yuji Adachi, Sachiko Nakamura, Yasuo Ariki, Tomoko Hayashi, Tetusya Takiguchi (Kobe University)

Paper ID: 32 - Propagated Perturbation of Adversarial Attack for well-known CNNs: Empirical Study and its Explanation
jihyeun yoon (LG CNS), kyungyul kim (LG CNS), jongseong jang (LG CNS)*

Paper ID: 33 - Attention Guided Metal Artifact Correction in MRI using Deep Neural Networks
Jee Won Kim (KAIST), Kinam Kwon (KAIST), Byungjai Kim (KAIST), HyunWook Park (KAIST)*

Paper ID: 35 - Free-Lunch Saliency via Attention in Atari Agents
Dmitry Nikulin (Samsung AI Center Moscow)*, Anastasia Ianina (Samsung AI Center Moscow), Vladimir A Aliev (Samsung AI Center, Moscow), Sergey I Nikolenko (PDMI RAS)

, Paper ID: 37 - Fooling Neural Network Interpretations via Adversarial Model Manipulation
Juyeon Heo (Sungkyunkwan University), Sunghwan Joo (Sungkyunkwan University), Taesup Moon (Sungkyunkwan University)*

Paper ID: 39 - Second-order feature representation for visualizing pyramidal multiscale superpixel pooling network
Ali Tousi (Ulsan National Institute of Science and Technology)*, Jaesik Choi (KAIST)

Paper ID: 42 - Interpreting Undesirable Pixels for Image Classification on Black-Box Models
Sin-Han Kang (Korea University)*, Hong-Gyu Jung (Korea University), Seong-Whan Lee (Korea University)

Contact

UNIST Explainable Artificial Intelligence Center
    Jaesik Choi / jaesik.choi@kaist.ac.kr
    Sohee Cho / sohee.cho@kaist.ac.kr
    GyeongEun Lee / socool@unist.ac.kr / 052-217-2196

Organizers

Jaesik Choi @UNIST
Seong-Whan Lee @Korea Univ.
K.-R. Müller @TU Berlin
Seongju Hwang @KAIST
Bohyung Han @SNU
David Bau @MIT
Ludwig Schubert @OpenAI
Yong Man Ro @KAIST