The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Invariance, Geometry and Deep Neural Networks with Pavan Turaga - #386
10 snips
Jun 25, 2020 
In this engaging conversation, Pavan Turaga, an Associate Professor at Arizona State University, shares his groundbreaking work at the intersection of physics and computer vision. He dives into the complexities of invariance and the geometric foundations of deep learning. Pavan highlights the challenges of modeling image variability for object recognition and the innovative use of time constraints in activity classification. His insights into robust loss functions and the integration of artistic elements in technology reveal a fresh perspective on the field.
AI Snips
Chapters
Transcript
Episode notes
Vision's Unique Challenge
- Computer vision is a unique field, distinct from other machine learning applications due to the variability of the natural world.
- This variability, stemming from physics-based factors like lighting and viewpoint, poses a challenge for traditional statistical methods.
Two Approaches to Vision
- The core problem in computer vision is handling the variability in how objects appear due to changing physics-based factors like lighting and viewpoint.
- Two approaches exist: data-driven methods that rely on large datasets and physics-based models that incorporate physical knowledge.
Invariance in Vision
- The concept of "invariance" is key to computer vision, representing the intrinsic properties of objects that remain constant despite changes in appearance.
- Achieving invariance can be approached through physics-based rendering or data-driven methods focusing on minimal necessary information.