| KEYWORD |
Multi-DNN Neural Architecture Search for Efficient Co-Optimization on Resource-Constrained Platforms
keywords DEEP NEURAL NETWORKS, EDGE COMPUTING, MULTI-DNN INFERENCE, NEURAL ARCHITECTURE SEARCH
Reference persons ALESSIO BURRELLO, DANIELE JAHIER PAGLIARI
External reference persons Matteo Risso
Research Groups ELECTRONIC DESIGN AUTOMATION - EDA
Thesis type EXPERIMENTAL
Description Nowadays, the range of possible tasks that need deep learning to be accomplished is expanding. Moreover, in domains such as autonomous drones, wearable health trackers, and augmented reality (AR) devices, the need for multiple DNNs to perform various interdependent tasks has grown. However, such DNNs need to be deployed on resource-constrained battery-operated embedded systems with limited computational power and memory, and they may have to respect some real-time requirements.
The candidate will explore a novel neural architecture search (NAS) framework for co-optimizing multiple deep neural networks (DNNs) to run efficiently on resource-constrained platforms. First, the candidate will consider the case where the main requirement is fitting the available memory. Then, the candidate will possibly extend the approach to both memory and latency requirements.
The following are possible areas of application:
- Wearable health trackers where multiple DNNs are employed to provide continuous monitoring of vital signs and activity; for instance, one model is used to estimate the heart rate, another analyzes the blood pressure, and a third recognizes human activity patterns.
- Autonomous drones rely on multiple DNNs to navigate, avoid obstacles, and track objects.
- AR glasses where one model might track environmental objects, another processes user gestures, and a third overlays contextual information based on the scene.
Required skills Proficiency in Python is required. Familiarity with Deep Learning, and PyTorch.
Deadline 31/10/2025
PROPONI LA TUA CANDIDATURA