Practical AI Attack of the C̶l̶o̶n̶e̶s̶ Text!
Aug 3, 2020
Jack Morris, a researcher at the University of Virginia and an incoming AI resident at Google, dives into the world of adversarial attacks in NLP. He sheds light on TextAttack, a framework that empowers developers to create stronger NLP models through adversarial examples and data augmentation. The conversation touches on the vulnerabilities of NLP models and discusses strategies to enhance their robustness. Cultural biases in training data and their impact on sentiment classification also spark an intriguing dialogue as they explore the future of AI understanding.
AI Snips
Chapters
Transcript
Episode notes
Google FUBAR
- Jack Morris was invited to interview at Google through an Easter egg found while searching about Python list comprehensions.
- After completing coding challenges, he received an interview and subsequent internship.
Adversarial Examples in NLP
- Adversarial examples research reveals model vulnerabilities, similar to computer vision.
- Adding imperceptible noise to images can cause misclassification, highlighting model limitations.
Sentiment Analysis Example
- Sentiment analysis models can be easily fooled by switching proper nouns.
- Changing "United States" to "Turkey" in a positive sentence might turn the sentiment negative, revealing biases.