Abstract
This chapter covers the issue of reducing the communication cost for machine learning at the edge from the perspective of data compression. Unlike traditional data compression schemes that aim at supporting the reconstruction of the original data, here the compression only needs to support the learning of the models that need to be learned from the original data, in order to support AI applications in a bandwidth-limited edge network. This lowered goal opens the door to a variety of application-specific lossy compression schemes designed to support machine learning. The focus in this chapter is on a subset of such schemes that can construct a weighted dataset much smaller than the original dataset that can function as a replacement of the original dataset in learning tasks, known as coreset. It reviews the history of coresets and their limitations, and then details two recently proposed improvements on (1) robust coreset construction and (2) integration of coreset construction and quantization.
Original language | English (US) |
---|---|
Title of host publication | Artificial Intelligence for Edge Computing |
Publisher | Springer International Publishing |
Pages | 223-264 |
Number of pages | 42 |
ISBN (Electronic) | 9783031407871 |
ISBN (Print) | 9783031407864 |
DOIs | |
State | Published - Dec 21 2023 |
All Science Journal Classification (ASJC) codes
- General Computer Science
- General Engineering