The NIPS '17 Competition: Building Intelligent Systems

Name: The NIPS '17 Competition: Building Intelligent Systems
Brand: Springer
Price: 53.49 EUR
Availability: OnlineOnly

Sergio Escalera Markus Weimer(Editor)

Springer (Publisher)

1st Edition

Published on 27. September 2018

X, 287 pages

E-Book

PDF with digital watermarking

System requirements

978-3-319-94042-7 (ISBN)

€53.49incl. 7% vat

System requirements

for PDF with digital watermarking

E-Book Single Licence

Available for download

Description

More details

Other editions

Content

Intro
Foreword
Contents
1 Introduction to NIPS 2017 Competition Track
1.1 The Conversational Intelligence Challenge
1.1.1 Task
1.1.2 Running the Competition
1.1.3 Outcomes
1.2 Classifying Clinically Actionable Genetic Mutations
1.2.1 Task
1.2.2 Data
1.2.3 Running the Competition
1.2.4 Outcomes
1.3 Learning to Run
1.3.1 Task
1.3.2 Running the Competition
1.3.3 Outcomes
1.4 Human-Computer Question Answering Competition
1.4.1 Data
1.4.1.1 Test Set
1.4.2 Competition
1.4.2.1 Machine Evaluation
1.4.2.2 Human-Machine Evaluation
1.4.3 Outcomes
1.5 Adversarial Attacks and Defenses
1.5.1 Task and Evaluation Metrics
1.5.2 Dataset
1.5.3 Running the Competition
1.5.4 Outcomes
1.6 IBM Watson AI XPRIZE Milestones
1.6.1 Running the Competition
1.6.2 Outcomes
1.7 Neural Art Challenge
1.7.1 Task and Evaluation Metric
1.7.2 Running the Competition
1.7.3 Outcomes
1.8 Discussion
References
2 The First Conversational Intelligence Challenge
2.1 Introduction
2.2 Competition Description
2.2.1 Task
2.2.2 Structure of the Competition
2.2.3 Pre-dialogue Texts
2.2.4 Evaluation
2.2.5 Volunteers
2.2.6 Framework
2.3 Participants
2.3.1 RLLChatbot
2.3.1.1 Generation of Candidate Responses Based on News Article and Dialogue History
2.3.1.2 Controlling the Conversation Flow
2.3.2 kAIb
2.3.3 bot#1337
2.3.4 Poetwannabe
2.3.5 PolyU
2.3.6 DeepTalkHawk
2.4 Results
2.4.1 First Round
2.4.2 Final Round
2.5 Dataset
2.6 Discussion and Future Work
2.6.1 Data
2.6.2 Evaluation
2.6.3 Organisational Aspects
2.6.4 Future Work
References
3 ConvAI Dataset of Topic-Oriented Human-to-Chatbot Dialogues
3.1 Introduction
3.2 Data Collection
3.3 Statistics of Dialogues
3.4 Dialogue-Level Evaluation
3.5 Utterance-Level Evaluation
3.6 Other Properties of Data
3.7 Conclusion
References
4 A Talker Ensemble: The University of Wroclaw's Entry to the NIPS 2017 Conversational Intelligence Challenge
4.1 Introduction
4.2 Background and Related Work
4.3 System Description
4.3.1 Detailed Response Generation Algorithm
4.3.1.1 Common Utterance Pre- and Post-processing
4.3.2 Detailed Talker Descriptions
4.3.2.1 Simple Wikipedia Talker (SWT)
4.3.2.2 Wikipedia QA
4.3.2.3 DBpedia Talker
4.3.2.4 Topic Guess Talker
4.3.2.5 K-NN Talker
4.3.2.6 Wikiquote Talker
4.3.2.7 Alice Talker
4.3.2.8 Simple Fact Generator
4.3.2.9 Trivia Questions
4.3.2.10 Miscellaneous Talkers (Abacus, Gimmick)
4.3.3 Balancing and Prioritizing Talker Confidences
4.4 Conclusions and Future Work
Appendix: Sample Dialogues
References
5 Multi-view Ensemble Classification for Clinically Actionable Genetic Mutations
5.1 Introduction
5.2 Preliminary
5.2.1 Notations
5.2.2 Validation Set
5.2.3 Evaluation Metric
5.3 The Proposed Approach
5.3.1 Document View
5.3.1.1 Domain Knowledge
5.3.1.2 Document-Level Feature
5.3.1.3 Sentence-Level Feature
5.3.1.4 Word-Level Feature
5.3.1.5 Dimension Reduction
5.3.2 Entity Text View
5.3.2.1 View Construction
5.3.2.2 Feature Generation
5.3.3 Entity Name View
5.3.3.1 Character-Level Feature
5.3.3.2 Word Embedding Feature
5.3.4 Classifiers
5.4 Model Ensembles
5.5 Experimental Results
5.5.1 Experimental Settings
5.5.2 Effects of Multi-view Features
5.5.3 Results of Basic Models
5.5.4 Results of Model Ensemble
5.6 Conclusion
References
6 Learning to Run Challenge: Synthesizing Physiologically Accurate Motion Using Deep Reinforcement Learning
6.1 Overview of the Competition
6.2 Prior Work
6.3 Competition Description
6.3.1 Background
6.3.2 OpenSim Simulator
6.3.3 Tasks and Application Scenarios
6.3.4 Baselines and Code Available
6.3.5 Metrics
6.4 Organizational Aspects
6.4.1 Protocol
6.4.2 Execution
6.4.3 Problems and Solutions
6.4.4 Submissions
6.5 Results
6.6 Discussion
Affiliations and Acknowledgments
Appendix
Installation
References
7 Learning to Run Challenge Solutions: Adapting Reinforcement Learning Methods for Neuromusculoskeletal Environments
7.1 Introduction
7.2 Learning to Run with Actor-Critic Ensemble
7.2.1 Methods
7.2.1.1 Dooming Actions Problem of DDPG
7.2.1.2 Inference-Time ACE
7.2.1.3 Training with ACE
7.2.2 Experiments and Results
7.2.2.1 Baseline Implementation
7.2.2.2 ACE Experiments
7.2.3 Discussion
7.3 Deep Deterministic Policy Gradient and Improvements
7.3.1 Methods
7.3.1.1 DDPG Improvements
7.3.1.2 Parameter Noise
7.3.1.3 Layer Norm
7.3.1.4 Actions and States Reflection Symmetry
7.3.2 Experiments and Results
7.3.2.1 Benchmarking Different Models
7.3.2.2 Testing Improvements of DDPG
7.3.3 Discussion
7.4 Asynchronous DDPG with Deep Residual Network for Learning to Run
7.4.1 Methods
7.4.1.1 Asynchronous DDPG
7.4.1.2 The Neural Network Structure
7.4.1.3 Noise for Exploration
7.4.1.4 Detailed Implementation
7.4.2 Experiments and Results
7.4.2.1 The Number of Parallel Process
7.4.2.2 The Neural Network Structure
7.4.3 Discussion
7.5 Proximal Policy Optimization with Policy Blending
7.5.1 Methods
7.5.1.1 Combining Cautious and Aggressive Strategies
7.5.1.2 Frameskip
7.5.1.3 Reward Shaping
7.5.1.4 Final Tuning: Policy Blending
7.5.2 Experiments and Results
7.6 Double Bootstrapped DDPG for Continuous Deep Reinforcement Learning
7.6.1 Methods
7.6.1.1 Double Bootstrapped DDPG
7.6.1.2 Observation Preprocessing
7.6.1.3 Noise Schedule
7.6.2 Experiments
7.6.2.1 Details
7.6.2.2 Results
7.6.3 Discussion
7.7 Plain DDPG with Soft Target Networks
7.7.1 Method
7.7.1.1 State Description
7.7.1.2 Training Process
7.7.2 Experiments and Results
7.8 PPO with Reward Shaping
7.8.1 Methods
7.8.1.1 Reward Shaping
7.8.1.2 Feature Engineering
7.8.1.3 Normalizing Observations
7.8.2 Experiments and Results
7.8.3 Discussion
7.9 Leveraging Emergent Muscles-Activation Patterns: From DDPG in a Continuous Action Space to DQN With a Set of Actions
7.9.1 Methods
7.9.2 Discussion
Affiliations and Acknowledgments
References
8 Reinforcement Learning to Run. Fast
8.1 Introduction
8.2 Proximal Policy Optimization
8.3 Policy Representation
8.4 Training Approach
8.4.1 Dealing with the Finite Horizon
8.4.2 Exploiting Task Symmetry
8.4.3 Staged Training Regime
Stage I: Global Policy Initialization
Stage II: Policy Refinement
Stage III: Policy Specialization
8.5 Results
Appendix: Hyperparameters
References
9 Human-Computer Question Answering: The Case for Quizbowl
9.1 What Is Quizbowl?
9.2 Why Should This Be a Computer Competition?
9.2.1 Who's Smarter?
9.2.2 Machine Learning
9.2.3 Natural Language Processing
9.3 A Simple System to Scaffold Users
9.3.1 Guesser
9.3.2 Buzzer
9.4 Client-Server Architecture
9.5 Results
9.5.1 Computer Results
9.5.1.1 Accuracy
9.5.1.2 Simulated Games
9.5.2 Human-Computer Games
9.5.2.1 First-Place Game
9.5.2.2 Second-Place Game
9.5.3 Are Computers Better at Quizbowl?
9.5.4 Improving Participation
9.6 Long-Term Viability of Increasingly Difficult Competitions
9.6.1 Changing the Difficulty
9.6.2 Speech Recognition
9.6.3 Adversarial Question Writing
9.6.4 Bonus Questions
9.6.5 Open Domain Answers
9.7 Comparison to Other Tasks
9.8 Conclusion
References
10 Studio Ousia's Quiz Bowl Question Answering System
10.1 Introduction
10.2 Proposed System
10.2.1 Data
10.2.2 Neural Quiz Solver
10.2.2.1 Model
10.2.2.2 Entity Detection
10.2.2.3 Pretrained Representations
10.2.2.4 Other Details
10.2.3 Neural Type Predictor
10.2.4 Information Retrieval Models
10.2.5 Answer Scorer
10.3 Experiments
10.3.1 Setup
10.3.2 Results
10.4 Competing with Other Systems and Human Experts
10.5 Conclusions
References
11 Adversarial Attacks and Defences Competition
11.1 Introduction
11.2 Adversarial Examples
11.2.1 Common Attack Scenarios
11.2.2 Attack Methods
11.2.2.1 White Box Digital Attacks
11.2.2.2 Black Box Attacks
11.2.3 Overview of Defenses
11.3 Adversarial Competition
11.3.1 Dataset
11.3.2 Tasks and Competition Rules
11.3.3 Evaluation Metrics
11.3.4 Competition Schedule
11.3.5 Technical Aspects of Evaluation
11.4 Competition Results
11.5 Top Scoring Submissions
11.5.1 First Place in Defense Track: Team TsAIL
11.5.1.1 Dataset
11.5.1.2 Denoising U-net
11.5.1.3 Loss Function
11.5.2 First Place in Both Attack Tracks: Team TsAIL
11.5.2.1 Method
11.5.2.2 Submission for Non-targeted Attack
11.5.2.3 Submission for Targeted Attack
11.5.3 Second Place in Defense Track: Team iyswim
11.5.3.1 Randomization as Defense
11.5.3.2 Randomization Layers
11.5.3.3 Randomization Layers + Adversarial Training
11.5.3.4 Submission Details and Results
11.5.3.5 Attackers with more Information
11.5.4 Second Place in both Attack Tracks: Team Sangxia
11.5.5 Third Place in Targeted Attack Track: Team FatFingers
11.5.5.1 Targeted Attack Model Transfer
11.5.5.2 Ensemble Attack Methods
11.5.5.3 Dynamic Iterative Ensemble Attack
11.5.6 Fourth Place in Defense Track: Team erko
11.5.6.1 Architecture of Defense Model
11.5.6.2 Spatial Smoothing: Median Filtering
11.5.6.3 Experiments
11.5.6.4 Effects of Median Filtering
11.5.6.5 Submission Results
11.5.7 Fourth Place in Non-targeted Attack Track: Team iwiwi
11.5.7.1 Basic Framework
11.5.7.2 Empirical Enhancement
11.5.7.3 Results and Discussion
11.6 Conclusion
References
12 First Year Results from the IBM Watson AI XPRIZE: Lessons for the ``AI for Good'' Movement
12.1 Introduction
12.2 Prize Process
12.3 Case Studies in the AI for Good Movement
12.3.1 Top Performing Problem Domains
12.3.2 Underperforming Problem Domains
12.3.3 Ethical Review for Artificial Intelligence Systems
12.4 Conclusions
References
13 Aifred Health, a Deep Learning Powered Clinical Decision Support System for Mental Health
13.1 Introduction: What Is Aifred Health?
13.2 Problem Description: What Is Depression and Why Is It Hard to Treat?
13.3 Solution Description and Rationale
13.4 Technical Description
13.5 Optimizing Physician-AI Interaction
13.5.1 Transparency
13.5.2 Adaptability
13.5.3 Integration
13.6 Clinical Validation
13.6.1 Ease of Use Study
13.6.2 Open Label Trial
13.6.3 Outcomes
13.6.4 Randomized Control Trial
13.7 AI Ethics: Meticulous Transparency
13.7.1 Details of an MT Assessment
13.8 Conclusion
References

System requirements

Save as PDF Copy link into clipboard

Schweitzer Fachinformationen

The NIPS '17 Competition: Building Intelligent Systems

Description

More details

Other editions

Additional editions

Content

System requirements