Keynotes

Abstract
Smartphones, wearable devices, and other computational units that are ubiquitous in our environments are imbued with increasingly more complex sensing, computational, and communication capabilities. These devices can generate (and distribute) vast quantities of data that can be used to build sophisticated machine learning models for a variety of applications, e.g., classification and recommendation. Opportunistic collaborative learning (OppCL) is a framework for individual devices in pervasive computing environments to train a deep learning model that caters to the device’s personalized needs. In OppCL, each device maintains a local, personalized model. When the device encounters another device via peer-to-peer communication, it shares its model parameters and asks the neighbor to train the model using the neighbor’s local data. This talk will present the motivation and use cases behind the creation of OppCL and a basic model for collaboratively training personalized models using opportunistically available neighboring devices (and their data!). The talk will discuss multiple schemes for incorporating encountered model updates as well as techniques for handling heterogeneity in the pervasive computing environment, including bandwidth and latency constrained communication links as well as computationally constrained neighboring devices. The talk will also include presentations of practical implementations for OppCL in both large scale simulation and in real world devices. The talk will close with a look forward into open challenges and opportunities in employing OppCL to diverse pervasive computing applications.