Doctoral Thesis: Continuous Learning for Lightweight Machine Learning Inference at the Edge
Abstract: Many applications require Deep Neural Network (DNN) inference on edge devices to deliver a robust performance or comply with privacy constraints. However, standard edge devices, like mobile phones, have significantly less compute power than cloud-based accelerators, and state-of-the-art DNN models are usually too expensive to run on such devices. On the other hand, lightweight neural network models that can run much faster on such devices at the edge typically have a significant accuracy gap with state-of-the-art.
In this talk, I will present my work on systems that use continuous learning to enable accurate and lightweight ML inference at the edge. In particular, I will discuss adaptive model streaming (AMS), a new method for remotely adapting a model deployed at the edge over the network with low communication overhead. I will describe applications of AMS in real-time video analytics and its use in SRVC, a lightweight learning-based video compression scheme. SRVC continuously adapts a lightweight super-resolution neural network to specific video content, and it is the first learned compression scheme that outperforms H.265 in its slow mode preset. Finally, I will discuss RECL, that combines model retraining and reuse to improve the response time and resource efficiency of continuous learning across edge devices at scale.
- Date: Friday, November 18
- Time: 2:30 pm - 4:00 pm
- Category: Thesis Defense
- Location: 32-G449
Additional Location Details:
Thesis Supervisor: Prof. Mohammad Alizadeh
The event will be a hybrid of in-person and online meetings. For those who prefer to attend online, please contact the speaker for the Zoom link.