Machine Learning Street Talk (MLST) #52 - Unadversarial Examples (Hadi Salman, MIT)
May 1, 2021
Hadi Salman, a PhD student at MIT with experience at Uber and Microsoft Research, dives into the intriguing world of adversarial and unadversarial examples. He discusses how slight image alterations can mislead classifiers and explores innovative ways to flip this problem on its head. By designing unadversarial examples, Hadi aims to create more robust models. The conversation also touches on the balance between accuracy and robustness, as well as the potential of adversarial training to enhance transfer learning outcomes.
AI Snips
Chapters
Transcript
Episode notes
Initial Fascination with Adversarial Examples
- Hadi Salman's interest in adversarial examples began with a talk at CMU.
- He was fascinated by how easily crafted examples could break well-performing models.
Worst-Case Performance of Models
- Adversarial examples revealed that models perform well on average but poorly in worst-case scenarios.
- Current machine learning models are far from robust and reliable for safety-critical applications.
Effective Research Practices
- Invest time in writing clear, well-structured papers with good code and blog posts to enhance impact.
- Prioritize problem selection carefully; spend time evaluating potential impact before starting.