IC Colloquium: Compositional Perception for Action
Event details
| Date | 04.04.2019 |
| Hour | 10:15 › 11:15 |
| Location | |
| Category | Conferences - Seminars |
By: Amir Zamir - Stanford University and UC Berkeley
IC Faculty candidate
Abstract:
Artificial Intelligence seeks agents that can perceive the world and act accordingly. Despite remarkable progress toward this goal, a fundamental shortcoming persists on the perception front: difficulty in scaling to the complexity of the real world, and consequently, reducing the operation domain to perceptually simplified ones (e.g. controlled spaces, video games, tabletop manipulation scenarios). I will present my efforts toward a visual perception that can ultimately scale to real-world complexity and support goals of active agents by going beyond isolated pattern recognition problems. The core idea is utilizing compositionality to tame the complexity: the world is largely composed of shared visual factors, hence a compact perceptual skill set can be sufficient for understanding large parts of it. I will show a method for tractably learning a large set of perception tasks using transfer learning (Taskonomy), toward forming a multi-task compositional perception dictionary. I will show this dictionary can be turned into a mid-level perception module for active robotic agents, enabling them to perceive in the real world and improve their sample efficiency and generalization. This is accomplished using both real robots as well as a virtual environment rooted in real spaces (Gibson Environment). I will conclude with discussing cross-task consistency and employing that as an intrinsic source of supervision for continual fine tuning of a compositional perception.
Bio:
Amir Zamir is a postdoctoral researcher at Stanford University and University of California, Berkeley. His research interests are broadly in computer vision and machine learning with a focus on transfer/self/un supervised learning and perception-for-robotics. He has been recognized with CVPR (2018) Best Paper Award, CVPR (2016) Best Student Paper Award, NVIDIA Pioneering Research Award (2018), and Stanford ICME Seed Award (2016), among others. His research has been covered by popular press outlets, such as NPR or The New York Times.
More information
IC Faculty candidate
Abstract:
Artificial Intelligence seeks agents that can perceive the world and act accordingly. Despite remarkable progress toward this goal, a fundamental shortcoming persists on the perception front: difficulty in scaling to the complexity of the real world, and consequently, reducing the operation domain to perceptually simplified ones (e.g. controlled spaces, video games, tabletop manipulation scenarios). I will present my efforts toward a visual perception that can ultimately scale to real-world complexity and support goals of active agents by going beyond isolated pattern recognition problems. The core idea is utilizing compositionality to tame the complexity: the world is largely composed of shared visual factors, hence a compact perceptual skill set can be sufficient for understanding large parts of it. I will show a method for tractably learning a large set of perception tasks using transfer learning (Taskonomy), toward forming a multi-task compositional perception dictionary. I will show this dictionary can be turned into a mid-level perception module for active robotic agents, enabling them to perceive in the real world and improve their sample efficiency and generalization. This is accomplished using both real robots as well as a virtual environment rooted in real spaces (Gibson Environment). I will conclude with discussing cross-task consistency and employing that as an intrinsic source of supervision for continual fine tuning of a compositional perception.
Bio:
Amir Zamir is a postdoctoral researcher at Stanford University and University of California, Berkeley. His research interests are broadly in computer vision and machine learning with a focus on transfer/self/un supervised learning and perception-for-robotics. He has been recognized with CVPR (2018) Best Paper Award, CVPR (2016) Best Student Paper Award, NVIDIA Pioneering Research Award (2018), and Stanford ICME Seed Award (2016), among others. His research has been covered by popular press outlets, such as NPR or The New York Times.
More information
Practical information
- General public
- Free
- This event is internal
Contact
- Host: Pascal Fua