PySAD
An open-source Python library that finds unusual events in streaming data, with models that learn from one sample at a time.
Selim Firat Yilmaz
Manager, Machine Learning ResearchAmerican Express, LondonPhD, Imperial College London
I build machine learning models that turn large, noisy data into compact, useful representations. I focus on foundation models for text, tabular, and time-series data streams, and on agentic algorithm discovery. My earlier work adds images, voice, and video for compression, generation, and prediction.
Finding unusual points in a live stream
Normal range
Unusual point
Selected work
Kaira
A PyTorch toolkit to build and test communication systems that use neural networks, from the channel model to the encoder and decoder.
Learning to interfere in multi-user image transmission
Several transmitters send images over the same channel resources at once. A multi-view autoencoder learns to use the interference, and the model grows to 64 users with only 2.4% more parameters to train.
Selim F. Yilmaz et al.
IEEE Transactions on Wireless Communications
Wireless image delivery with diffusion models
A diffusion model at the receiver repairs images that arrive over noisy wireless channels.
Selim F. Yilmaz et al.
IEEE INFOCOM workshops
Work done at Huawei R&D
Sentiment analysis in 100 languages
One model for 100 languages that gives each text one or more sentiment labels and gives rare labels more weight in training.
Selim F. Yilmaz et al.
IEEE Transactions on Neural Networks and Learning Systems
Private collaborative edge inference
Edge devices send their predictions together in one wireless transmission, which saves bandwidth and keeps each model private.
Selim F. Yilmaz et al.
IEEE Transactions on Machine Learning in Communications and Networking
Education
- PhD, Electrical and Electronic EngineeringImperial College LondonFocus: machine learning for communication and compression
- MS, Electrical and Electronic EngineeringBilkent University
- BS, Computer ScienceBilkent University