Thread by @SigProcessing

[Thread on #MachineLearning from #streaming #data]

Announcing our (@matthewnokleby, @haroonraja86, and myself) paper on "Scaling-up Distributed Processing of Data Streams for Machine Learning", which has been accepted by @ProceedingsIEEE (Preprint: arxiv.org/abs/2005.08854).

1/

It focuses on training of models from fast streaming data, with "fast" referring to the inability of a single machine to process each data sample in time before the next one arrives. Distributed training can help deal with this, but how many nodes, what minibatch size, etc.?

2/

The paper addresses these questions and shows that there is an appropriate regime for number of nodes and minibatch size per node where training can be near-optimal in terms of excess risk. Outside this regime, distributed processing / minibatching slows down learning.

3/

The regime is determined by the streaming rate of data, the communications speeds of internode links (e.g., InfiniBand interconnects or ethernet links), and the processing capabilities of individual machines.

4/

The focus of the paper is on convex objective functions as well as the nonconvex PCA problem. The ideas should work for #federatedlearning also. It is written in an overview style and should have something in it for both experienced and new researchers.

5/

The paper also discusses open Q's in relationship to the interplay between loss functions, streaming rate, and compute and communication speeds. We are hopeful the paper would be useful in advancing the research on distributed ML training.

Comments welcome.

6/6 [END OF THREAD]

@threadreaderapp unroll