This fully funded PhD project will develop efficient and reliable test-time adaptation methods that allow unimodal and multimodal foundation models to identify environmental changes and adapt during deployment. The central challenge is to improve robustness without relying on costly cloud-based retraining or violating the strict latency, memory and energy constraints of edge devices.
Research objectives
The successful candidate will investigate:
- Autonomous monitoring: Methods for detecting distribution shifts and quantifying model uncertainty across heterogeneous data streams.
- On-the-fly adaptation: Lightweight algorithms that can update or recalibrate models during inference without access to labelled target-domain data.
- Efficient foundation models: Parameter-efficient adaptation, model compression and resource-aware architectures suitable for edge hardware.
- Reliability under dynamic conditions: Methods for balancing adaptation accuracy, computational cost, energy consumption and real-time execution requirements.
The research will combine theoretical and algorithmic development with experimental validation using modern computer-vision and multimodal foundation models on edge-computing testbeds.
Research environment
The candidate will join the Adaptive & Agentic AI (A3) Lab in the Department of Electrical and Computer Engineering at Aarhus University. The project will be supervised by Associate Professor Behzad Bozorgtabar and co-supervised by Professor Qi Zhang.
The candidate will work in an international and collaborative research environment and will be encouraged to publish at leading machine-learning and computer-vision venues such as NeurIPS, ICLR and CVPR.
Candidate profile
Applicants should hold, or expect to obtain, a Master’s degree equivalent to 120 ECTS in computer science, computer engineering, electrical engineering, machine learning or a closely related quantitative discipline.
Strong proficiency in Python and deep-learning frameworks such as PyTorch is expected. Experience in one or more of the following areas is particularly relevant:
- Machine learning and computer vision
- Test-time or domain adaptation
- Foundation models and multimodal learning
- Model compression or parameter-efficient fine-tuning
- Edge AI and resource-efficient inference
Application deadline: 15 August 2026 at 23:59 CEST.
