- Two DCSL-ers graduate with their PhDs. Ran Xu, er Dr. Ran Xu, is now headed to Santa Clara for working at NVIDIA. His thesis work was titled “Approximation for Streaming Video Analytics on Mobile Devices“. Dr. Edgardo Barsallo is now joining DCSL as a Post-doctoral Research Associate. His work was titled “Reliability of Mobile Wearable OS”. This led to two security advisories from Google, and corresponding two patches. Most recently his work has led to a CVE being issued, on security of Android TVs, CVE-2021-0889.
- Saurabh was invited to Google’s Workshop on Federated Learning and Analytics. It was a great three-day event where academics and Googlers discussed the latest developments, what application scenarios are coming up for federated learning, and laid down some markers for near-term research and development. Saurabh’s talk was titled “Federated Learning at the Edge“. His talk as well as a treasure trove of short talks at the event is at:
https://events.withgoogle.com/2021-workshop-on-federated-learning-and-analytics/pre-recorded-talks/#content
Abstract
FL at the edge is a different way of approaching federated training that can deal with wimpy IoT nodes and slightly more powerful GPU-equipped edge nodes. Such solutions have to deal with vast heterogeneity in compute capability plus dynamically changing network conditions. Next, the devices can be compromised, through remote means or even physical tampering. Finally, the model that is learned must be simple enough to fit inside the devices and to perform real-time inferencing. We present the initial few attempts being made by a group of researchers (including us) to address these problems. If we are successful, we can bring FL to the IoT world, which offers the promise of unbounded number of devices that can collectively learn and execute models.