Project Description:

The proliferation of connected devices capable of rapid data gathering and processing is motivating the deployment of advanced machine learning (ML) techniques throughout networked systems. Recent methodologies for distributing ML model training through such systems have proposed learning architectures consisting of local, device-level model updates combined with global, system-level model aggregations. Pushing computation to the network edge introduces a new set of challenges due to the heterogeneous nature of edge networks and the data-intensive, latency-sensitive nature of contemporary ML applications. Specifically, IoT devices are often heterogeneous in (i) their available on-board communication and computation resources and (ii) the statistical properties of their local datasets and modeling objectives. This can lead to long training delays, high resource utilization, and a mismatch between global models and rapidly evolving local system properties.

To address these challenges, the objective of this project is to establish the foundation for context-aware distributed learning, a new distributed training paradigm that will adapt to the different dimensions of heterogeneity in intelligent networked systems. Research components to be investigated include the introduction of controlled local coordination among devices, facilitated by direct device-to-device (D2D) communications in contemporary wireless protocols, to complement the local update/global aggregation framework of contemporary distributed ML.

Start Date:

March 2023

Postdoc Qualifications:

The postdoc researcher should hold a PhD in electrical/computer engineering, computer science, industrial engineering, or equivalent with a background in distributed optimization and machine learning. The researcher should have a record of publications in the top journals and/or conferences for these areas.

Co-Advisors:

Christopher G. Brinton
cgb@purdue.edu
Electrical and Computer Engineering
www.cbrinton.net

Vaneet Aggarwal
vaneet@purdue.edu
Industrial Engineering
https://web.ics.purdue.edu/~vaneet/

Bibliography:

Christopher G. Brinton
cgb@purdue.edu
Electrical and Computer Engineering
www.cbrinton.net

Vaneet Aggarwal
vaneet@purdue.edu
Industrial Engineering
https://web.ics.purdue.edu/~vaneet/