Valérian Rey

Research project on multi-objective optimization with an EPFL PhD student.

• Development of theJacobian Descent(JD) algorithm, a generalization of gradient descent.

• Development ofTorchJD**: a library extending autograd for JD ( **310+ GitHub stars).

• Deep learning experiments with JD on vision datasets.

• Publication ofJacobian Descent For Multi-Objective Optimization.

• Development of theautogramengine to make JD less memory-intensive.

• Accepted to the PyTorch Conference EU to present the project:

Development of machine learning algorithms for connected objects: literature review, data collection and processing, exploration, model development, C implementation, maintenance and monitoring. Some projects:

• Blood pressure monitoring via deep learning models based on optical sensor data (discontinued).

• Activity classification algorithm from accelerometer data (in production).

• Real-time accelerometer calibration algorithm from user data (in production).

• Biomarker regression from urine test strip images (in production).

• Mentoring an intern (then junior) deep learning engineer for 9 months.

Research project onfederated learningfor malware detection from connected object traffic data.

Supervised binary classification with neural networks, unsupervised anomaly detection via autoencoders.

Use of federated learning as a privacy-preserving approach while leveraging more data sources in model training.

Publication ofFederated learning for malware detection in IoT devicesin the journal Computer Networks (400+ citations, best paper award).