Master Thesis: Value-of-Information-Based Data Selection for Multimodal Edge Intelligence

Join our Team

About this opportunity:

We are seeking a motivated Master's student to investigate intelligent sensing and data-selection strategies for multimodal XR devices operating under computational, energy, and wireless communication constraints.

The thesis will study how an edge device equipped with multiple sensors can determine what information should be sensed, locally processed, and transmitted to an edge server. Multimodal devices may collect data from RGB or stereo cameras, infrared or thermal cameras, depth or LiDAR sensors, and inertial sensors. Continuously processing and transmitting all measurements may lead to excessive bandwidth usage, energy consumption, and latency.

In many situations, only a subset of sensor observations is necessary for tasks such as localization, obstacle detection, depth estimation, scene understanding, or 3D reconstruction. You will develop an AI-assisted decision-making framework that estimates the value, relevance, uncertainty, and freshness of sensor information. Based on these factors and current device and network conditions, the system will select the most useful sensor modality, spatial region, temporal sample, or intermediate representation for transmission.

What you will do:

Review active sensing, value-of-information estimation, sensor scheduling, task-oriented communication, and resource-aware edge intelligence methods.

Formulate adaptive multimodal data selection for XR devices.

Develop a lightweight decision-making module that selects sensors, data regions, features, or time samples according to the task, sensor confidence, scene dynamics, device resources, and wireless conditions.

Investigate AI and machine-learning strategies, such as uncertainty-aware models, contextual bandits, reinforcement learning, or Bayesian decision-making, while ensuring that the decision process remains feasible on a resource-constrained device.

Design an edge-side processing pipeline that receives selectively transmitted multimodal observations and performs the required perception or reconstruction task.

Evaluate the approach against full multimodal transmission, fixed sensor-selection schemes, and conventional adaptive strategies.

Analyse task performance, bandwidth usage, energy consumption, latency, robustness, and information freshness.

The skills you bring:

Enrolled in or admitted to a Master's programme in Computer Science, Electrical Engineering, Computer Engineering, Robotics.

Basic knowledge of machine learning, computer vision, wireless communications, signal processing, or edge computing.

Programming experience in Python and familiarity with common machine-learning frameworks.

Interested in theoretical modelling, algorithm design, and experimental evaluation.

Analyse system-level trade-offs and interpret quantitative results.

The following are considered a plus:

Experience with reinforcement learning, contextual bandits, Bayesian inference, uncertainty estimation

Knowledge of multimodal perception, sensor fusion, semantic communication, or task-oriented communication.

Familiarity with PyTorch, TensorFlow, OpenCV, ROS/ROS 2, Open3D, or similar tools.

Experience with embedded devices, edge computing, wireless networks, or hardware profiling.