Learning and Adaptation in distributed environments
My research focuses on the investigation of new solutions to personalize and improve deep learning models for a specific client. In particular, I focus on time-series data coming from MEMS sensors, which include accelerometers and gyroscopes. To pose my research in a more realistic scenario, I deal with real-world applications where data are sensitive and cannot be shared to the main server. On top of that, due to the lack of labels on clients' data, I'm exploring Semi-Supervised solutions with the aim to extract information from clients' data in an unsupervised way.
More in detail, I've published a paper on IEEE Embedded System Letters which deonstrate the importance of personalization in a HAR scenario, due to the heterogeneity of subjects. In this feasibility study we show that a simple 1D-CNN model can be fully retrained on-device and that this personalization step improve significantly the accuracy of the model. Moreover, starting from this work, I'm exploring the world of Semi-Supervised Federated Learning and the integration of lightweight Domain Adaptation solutions within this popular paradigm.
In addition to that, I'm currently working on a collaboration with the biomedical department to solve a segmentation problem on the seismocardiogram (SCG) signal. SCG is a mechanical signal measured with accelerometers and gyroscopes, which measures the micro-movement of the chest caused by heartbeats. In particular, I've constructed a fully convolutional network which automatically extract relevant features from portions of raw SCG and segments the systolic complex of every heartbeat. Thanks to this segmentation we can derive the heart rate variability directly from the SCG and use this mechanical signal as substitute of the classical ECG.
Back to Current Students