SplitEasy: A Practical Approach for Training ML models on Mobile Devices
- Position: Machine Learning Research Engineer
- Type: Part-Time
- Duration: June 2020 - December 2020
- Supervisor: Dr. Dimitris Chatzopoulos
- Location: Hong Kong University of Science and Technology, Kowloon, Hong Kong
- Status: Published in ACM HotMobile 2021. An international workshop with ~30% acceptance rate.
- Visit: View Publication
Working as an assistant researcher and engineer to formulate and implement a computational offloading protocol that can support DL model training on edge devices with constant computation cost. Tweaked the backpropogation algorithm to train sensitive (first and last) layers on device, and the computationally intensive part on the server. Simulated the new algorithm on PyTroch and obtained benchmark results on MLP and CNN without exposing client data and labels. Implemented the protocol on a real mobile app (using Tensorflow.js) and currently working to measure the associated hardware and communication costs.