“A Day Inside an iGaming Hub” | 4 Ιουνίου 2026 (11:00–13:00)
May 12, 2026
The Department of Civil and Environmental Engineering announces one (1) open position at the rank of Lecturer or Assistant Professor in the field of: “Resource Sustainability Engineering”
May 12, 2026
Dynamic Early-Exit Convolutional Neural Networks for Edge Vision: Do only what and when you need!
by Theocharis Theocharides
May 21, 2026
4:00 PM (Athens Time)
Description
Computer vision has experienced remarkable growth in recent years, driven largely by advances in deep convolutional neural networks (CNNs). However, deploying these models on edge devices—such as autonomous drones and terrestrial robotic platforms—remains challenging due to limitations in power, energy consumption, computational resources, memory footprint, and real-time performance requirements.
These challenges become even more critical in safety-sensitive applications such as search and rescue missions and emergency management, where systems must operate reliably under difficult visual conditions including occlusions, changing environments, and dynamic operational contexts.
Traditional approaches focus on training CNNs to generalize across these conditions while optimizing them for embedded hardware platforms and accelerators. Although techniques such as pruning and quantization help compress and optimize deep neural networks, they are still constrained by the computational capabilities of the host device.
This lecture presents recent advances in dynamic deep convolutional neural networks for low-power embedded vision systems. Dynamic DNNs improve efficiency by adapting computation according to input complexity. Unlike static CNNs that process every input with fixed depth and width, dynamic architectures selectively activate layers, channels, or early exits, significantly reducing unnecessary computation and memory access.
This conditional execution enables:
- Lower energy consumption
- Faster inference
- Improved efficiency for real-time edge applications
Dynamic CNNs are therefore highly suitable for resource-constrained embedded systems and intelligent edge devices.
Join the next IEEE CEDA Virtual Distinguished Lecturer session.
21 May 2026 | 9:00 AM ET
Register now:
IEEE CEDA Registration Page
