CONCEPT DRIFT AND MODEL DEGRADATION FOR IOT DATA IN FOUNDATION MODELS
Introduction
Foundation models are becoming increasingly relevant for IoT and cyber-physical systems because they can learn from large-scale, heterogeneous, and multimodal sensor data and transfer across tasks more effectively than traditional task-specific models. However, IoT environments are inherently dynamic: sensors, users, contexts, and environments change over time, which can cause concept drift and gradual model degradation. Monitoring these effects is therefore essential for reliable long-term deployment of foundation models in IoT settings.
Objectives
· Study concept drift and model degradation for IoT data processed by foundation models.
· Implement and compare methods to detect changes in data distributions and model behavior over time.
· Evaluate when adaptation, retraining, or human intervention should be triggered.
Tasks
1. Literature Review: Review foundation models for IoT, concept drift, and monitoring of adaptive IoT/edge ML systems.
2. Benchmark Selection: Choose one or two public IoT/wearable/smart-environment datasets with temporal structure, and a suitable pretrained or foundation-model-based architecture.
3. Implementation: Build an experimental pipeline to evaluate the model under temporally evolving data conditions.
4. Drift and Degradation Analysis: Implement several indicators for data drift, feature drift, confidence changes, or downstream performance degradation.
5. Evaluation: Compare detection strategies in terms of sensitivity, false alarms, detection delay, computational cost, and usefulness for triggering model updates.
6. Discussion: Derive practical recommendations for when and how to trigger adaptation and reliable monitoring of foundation models in IoT applications.
Pre-requisites
Python, machine learning, Interest in IoT, time series, or foundation models is a plus.
Work
30% Theory, 50% Programming/Experiments, 20% Writing
Contact
Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)