Explore the cutting-edge field of natural language understanding through adversarial testing in this 59-minute Stanford University webinar. Delve into Professor Christopher Potts' insights on the current state and future potential of AI research, with a focus on his course XCS224U. Discover how adversarial testing challenges top-performing systems to reveal their weaknesses and drive innovation. Gain valuable perspectives from former students as they present original projects developed during the course, showcasing the practical application of concepts learned. Examine topics such as image captioning, natural language inference, and the evolution of AI systems like Watson. Learn about standard and adversarial evaluations, assignment structures, and the process of developing research projects from literature review to conclusion. Acquire a comprehensive understanding of the motivations, methodologies, and potential breakthroughs in natural language understanding.
Improving Natural Language Understanding Through Adversarial Testing
Stanford University via YouTube
Overview
Syllabus
Intro
Overview
Image captioning
Watson wins Jeopardy (2011)
Natural Language Inference (NLI)
Stanford Natural Language Inference (SNLI)
"Superhuman" performance on other tasks
The promise of artificial assistants
Two perspectives
Standard evaluations
Adversarial evaluations
NLI adversarial testing
ROBERTa evaluation
Adversarial NLI
High-level summary
Assignments and bake-offs
Assign/Bake-off: Word-level entailment
Assign/Bake-off: Contextual color describers
Wrap-up
Key Motivations
Literature Review to Project
Experiment Flow
Results and Analysis
Conclusions
Lessons Learned
Thank you
Taught by
Stanford Online
