Flower – A Friendly Federated Learning Framework | Hacker News Reader

Flower – A Friendly Federated Learning Framework

github.com

51 pointstanto13 comments

What is a 'federated learning framework'?

At first I thought this was something a Leaning Management System like Moodle, maybe one that is distributed sort of like a torrent so it can't be taken down.

Seeing the example talk about TensorFlow, it must have something to do with machine learning.

Correct. Federated Learning is basically Machine Learning but the Learning occurs directly on the same devices where the data is (e.g. phones). Each device trains a model and sends it to one central model, which combines all of them together. With this it is not necessary for the central model to know about the data which is good with regards to data protection laws.

For those interested or confused around federated learning, I suggest skimming the 2019 paper "Advances and Open Problems in Federated Learning" [0]. The introduction is quite good and the table of contents gives a good overview of the current challenges in the field.

From the paper, "Federated learning is a machine learning setting where multiple entities (clients) collaborate in solving a machine learning problem, under the coordination of a central server or service provider. Each client’s raw data is stored locally and not exchanged or transferred; instead, focused updates intended for immediate aggregation are used to achieve the learning objective."

0: https://arxiv.org/abs/1912.04977

Hey! I am one of the creators of Flower. If you have any questions regarding the framework we are happy to answer any. If you would like to dive a bit deeper and understand whats this about we have a great blog post [0] which gives a short intro into FL and motivates FL on embedded devices with an actual use-case.

0: https://flower.dev/blog/2020-12-16-running_federated_learnin...

Hey! Hope someone can help me with this question. How does the data exploration phase look like in a federated learning context? Most of the times (most probably always) before applying any ML algorithm we'd look at the data and explore it. How can this be done in this case if data is not available to see? Even in the example in the blog post of Flower the dataset is loaded directly without any pre-processing (which is usually the case in real life).

You can’t explore the data. Or rather you have to do it on a different dataset that you collect the traditional way. Federated learning is hard.

Thanks! Based on your own experience is that how you handled it? How did it work out in your case?

In my experience federated learning gets talked about a lot more than actually used.

I am not sure this is true. Do you have any numbers/data backing this?

I'm researching federated learning. It's currently used in a number of contexts including the Google and Apple keyboards on your Android and iOS devices respectively.

Federated learning is a very active field of research. There are no simple frameworks that folks can easily operationalize. Most do not have problems that necessitate federated learning—although the growth in data privacy laws, public-private partnerships, and need to build models on privately held data (think commercial partnerships) are making it more and more prevalent.

That's very interesting. What is the focus of your research in FL?

I am studying aspects of compression (i.e., gradient compression) in federated learning. I also study problems and applications of federated learning to public-private partnerships (i.e., the cross-silo setting as opposed to the cross-domain setting).

When I read "Friendly Federated Learning Framework", my immediate thought was something where people could gather and organize their personal knowledge, and have the ability to simply reference work that others have already done.

It's the internet.

Reply on news.ycombinator.com