The Thesis Review [32] Andre Martins - The Geometry of Constrained Structured Prediction
Sep 16, 2021
In this episode, Andre Martins, an Associate Professor at IST and VP of AI Research at Unbabel, dives into his pioneering work in natural language processing and structured prediction. He reflects on his PhD journey and the philosophical implications of AI in communication. The discussion highlights the balance between theoretical foundations and empirical methods, the evolution of sparse learning in neural networks, and advancements like SparseMax. Andre also emphasizes the importance of collaboration between academia and industry, offering insights into navigating a successful research career.
AI Snips
Chapters
Books
Transcript
Episode notes
Theory's Crucial Role in NLP
- Andre Martins believes theory should play a bigger role in NLP alongside empirical methods.
- Understanding assumptions and theoretical gaps is crucial despite neural networks' theoretical limits.
Early Path to Machine Learning
- Andre's first machine learning exposure came from his undergraduate thesis on computer vision.
- He shifted to NLP after industry work, drawn by his interests in math and linguistics.
Four Advisors and Shifting Focus
- Andre had four advisors from diverse backgrounds, which helped him explore various research directions.
- He transitioned from kernel methods to structured prediction focused on parsing.