4th Invited talk - LINCS Annual Workshop with its Scientific Commitee

“A brief (machine learning) foray at the edge of computing”, by Roch Guerin (Washington University in Saint Louis).

Edge computing solutions have proliferated, fueled by a combination of increased network ubiquity, advances in computing, especially in embedded devices, and by the growing need to bring computations closer to where data is produced. Many of those scenarios are driven by machine learning applications. In this talk, I will discuss two projects, both motivated by edge computing machine learning applications, and for which machine learning was itself instrumental in devising an efficient solution.

1. The first project [1] targets object detection with local and edge compute resources cooperating to optimize detection accuracy under load constraint on the edge server. Under such a constraint, the goal is to devise a simple policy to decide which images to offload to the edge server while maximizing detection accuracy. This calls for a metric that quantifies improvements in overall detection accuracy from offloading an individual image, and an estimator for that metric that can run on embedded devices. The benefits of the approach are again demonstrated experimentally.

2. The second project deals with an object classification problem where a camera is uploading images to an edge server for classification [2]. The wireless network used to upload successive images is, however, subject to bandwidth fluctuations. This requires an adaptive transmission strategy to maximize inference accuracy, irrespective of the amount of data that can be transmitted for each image. We realize this through a simple application of stochastic tail-drop when training a neural compression algorithm and demonstrate the efficacy of the approach on a local testbed.

[1] J. Qiu, R. Wang, B. Hu, R. Guerin, and C. Lu, “Optimizing Edge Offloading Decisions for Object Detection.” Under submission, 2024.

[2] R. Wang, H. Liu, J. Qiu, M. Xu, R. Guerin, and C. Lu, "Progressive Neural Compression for Adaptive Image Offloading Under Timing Constraints." Best Student Paper Award, 2023 IEEE Real-Time Systems Symposium (RTSS), December 2023, Taipei, Taiwan.