Perception-Action Cycle: Implications of Active Learning for Developing Visual Intelligence
Event details
| Date | 24.08.2021 |
| Hour | 15:30 › 17:30 |
| Speaker | Onur Beker |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Volkan Cevher
thesis advisor: Prof. Amir Zamir
co-examiner: Prof. Mackenzie Mathis
Abstract
Even simple animals like insects, fish, or frogs have a remarkable visual capacity to robustly navigate, interact with, and survive in the real world. And although there is a significant body of experimental work aimed at discovering the general principles underlying these capacities, it is yet unclear how the brain decodes visual stimuli to high-level representations (e.g. objects and events) that can guide decisions and actions. Similarly, the standard task-centric approaches in computer vision (e.g. detecting objects, computing optic flow or building depth maps), while definitely useful, are not yet sufficient to enable autonomy and adaptive behavior in complex real-world environments, at a level of flexibility comparable to that of natural intelligence. An alternative direction is to instead adopt an agent-centric approach to the problem, and conduct an exploratory computational study of active vision by iteratively optimizing autonomous systems on down-stream tasks that require a non-trivial visual capacity for successful completion, and analyzing the implications of this optimization procedure on the perception models of the said systems. This way, neither vision nor active behavior is treated as a stand-alone function, and the relationship between them is instead formulated as that of co-dependent participants with reciprocal contributions in a perception-action cycle, rather then that of a benefactor serving a beneficiary.
Background papers
exam president: Prof. Volkan Cevher
thesis advisor: Prof. Amir Zamir
co-examiner: Prof. Mackenzie Mathis
Abstract
Even simple animals like insects, fish, or frogs have a remarkable visual capacity to robustly navigate, interact with, and survive in the real world. And although there is a significant body of experimental work aimed at discovering the general principles underlying these capacities, it is yet unclear how the brain decodes visual stimuli to high-level representations (e.g. objects and events) that can guide decisions and actions. Similarly, the standard task-centric approaches in computer vision (e.g. detecting objects, computing optic flow or building depth maps), while definitely useful, are not yet sufficient to enable autonomy and adaptive behavior in complex real-world environments, at a level of flexibility comparable to that of natural intelligence. An alternative direction is to instead adopt an agent-centric approach to the problem, and conduct an exploratory computational study of active vision by iteratively optimizing autonomous systems on down-stream tasks that require a non-trivial visual capacity for successful completion, and analyzing the implications of this optimization procedure on the perception models of the said systems. This way, neither vision nor active behavior is treated as a stand-alone function, and the relationship between them is instead formulated as that of co-dependent participants with reciprocal contributions in a perception-action cycle, rather then that of a benefactor serving a beneficiary.
Background papers
Practical information
- General public
- Free
Contact
- edic@epfl.ch