Closed
Bug 1462102
Opened 6 years ago
Closed 5 years ago
[meta] Federated Learning
Categories
(Data Platform and Tools :: General, enhancement, P3)
Data Platform and Tools
General
Tracking
(Not tracked)
RESOLVED
FIXED
People
(Reporter: fhartmann, Assigned: fhartmann, Mentored)
References
Details
Attachments
(1 file)
5.75 KB,
text/plain
|
Details |
Federated Learning [1, 2] is a new subarea of machine learning that allows training models without directly collecting training data. Instead, clients improve the global model by locally training using their private data and then uploading weight updates. We want to experiment with this idea and use it to train a model for Firefox. [1] https://ai.googleblog.com/2017/04/federated-learning-collaborative.html [2] https://florian.github.io/federated-learning/
Updated•6 years ago
|
Priority: -- → P3
Assignee | ||
Comment 1•6 years ago
|
||
We decided to use federated learning to optimize the history suggestions in the awesome bar. Currently they are ranked using the frecency algorithm [1], which assigns scores to every URL based on how recently and frequently it was visited. Federated learning can be used to optimize the weights used in that algorithm. For example, frecency currently adds a score of 100 if the website was visited during the last four days. These constants (100, 4) weren't chosen in a data-driven process, so the idea is that they can still be optimized using federated learning. A first implementation of this for the client part is available on GitHub [2]. [1] https://developer.mozilla.org/en-US/docs/Mozilla/Tech/Places/Frecency_algorithm [2] https://github.com/florian/federated-learning-addon
Assignee | ||
Comment 2•6 years ago
|
||
The addon is now fully functional. It is currently being reviewed on GitHub [1]. [1] https://github.com/florian/federated-learning-addon/pull/1
Assignee | ||
Comment 3•6 years ago
|
||
Assignee | ||
Comment 4•6 years ago
|
||
We launched our SHIELD study two days ago. On the first day, only the client-side part was deployed. Yesterday, we enabled the server-side part to start the optimization process. The first results are already coming in and look pretty promising so far.
What did you do? Firefox just reset itself (erasing all my Addons and Bookmarks) and only some "Federated Learning" addon was left!!
(In reply to James from comment #5) > What did you do? Firefox just reset itself (erasing all my Addons and > Bookmarks) and only some "Federated Learning" addon was left!! Hi James, that sounds like you've experienced a profile reset and it's just a coincidence that you were enrolled in the federated learning study afterwards. You can disable or uninstall the addon if you'd like. If you continue to experience problems let us know and we'll look into it, but this is the first report we've heard about an issue like this so the study is most likely not the issue here.
Comment 7•6 years ago
|
||
I'll add that I've recently experienced weird behavior around addons installed. I can't isolate it, but I lost one of my addons that was installed seemingly when this one was installed.
Comment 8•6 years ago
|
||
To give a short update on this: I published a blog post that provides an overview of the project and shows some of the results: https://florian.github.io/federated-learning-firefox/ The optimization process itself worked very well, as the loss kept decreasing over time. Users in the treatment group type roughly half a character less to find what they are looking for. However, they also select suggestions that are not as far on top of the list as in the control groups. In case this is not desirable, it would only require very minor changes in the existing system to optimize for something different.
Updated•5 years ago
|
Status: NEW → RESOLVED
Closed: 5 years ago
Resolution: --- → FIXED
You need to log in
before you can comment on or make changes to this bug.
Description
•