EMBEDDED LEARNING�AND SENSING SYSTEMS (ELSS)
The Embedded Learning and Sensing Systems (ELSS) group works on the development of AI-powered embedded systems. Modern computational models used to process IoT sensor data are increasingly based on deep learning principles. These models exert severe demands on the local resources, need to adapt to the target context, and ensure robustness to unexpected changes. We believe in "less is more" and "small can be mighty" when it comes to running deep models on resource-constrained devices.
- Resource-efficient machine learning
- Deep network compression
- Weight-space ensembles of deep neural networks
- Distributed learning and optimization
- On-device model training & adaptation
- On-device learning, parameter-efficient fine-tuning
- Domain adaptation and model reconfiguration
- Mitigating catastrophic forgetting in continual learning
- Engineering robust & safe AI-based sensing systems
- Multi-modal sensing and multi-modal learning
- Adversarial attacks and adversarial patches
- Interpretability of deep models
Institute of Technical Informatics
Inffeldgasse 16/I, 8010 Graz, Austria
office@iti.tugraz.at
Efficient deep ensembles & where to find them
- Can deep nets be constructively merged in the weight space?
- No additional resources required at inference time
- Support for multi-task learning and distributed training
- We show that weight-�space ensembles of �SGD solutions are �possible modulo per-�mutation invariance
Distribution shifts & robustness
- AI-based sensing systems often process field data that is different from the training set, e.g., in orientation or lighting
- How to build robust models capable of generalizing to natural distribution shifts?
- We solve the problem �in the context of �pollen sensing from �microscopic images
Adaptation to dynamic resource constraints
- Machine learning pipelines generate resource-agnostic models
- How to adapt model inference to dynamically changing resource availability at runtime?
- We train a nested �model structure with�each sub-model �being effective for �its resource class