Hi all,
I am a PhD student working on Federated Learning (i.e. a decentralised learning where several parties collaboratively train ML models without revealing their data). More specifically, I am focusing on security and scalability concerns.
As I am starting my research, I am also looking forward to new applications and, as a Fediverse enthusiast, I would love to contribute to some of your projects. So I would like to know whether you have some ML tasks that you would like to integrate in one of the Fediverse project you are working on? My intent is not to force ML into Fediverse but simply to open this discussion.
The federation architecture is really interesting from a technical AND scientific POV: It creates new trust model in between the fully decentralised and the centralised setups. Moreover, these federated services have some data on the servers and on the users. Individually, they could be too small to train a decent ML model, but all together, they could be a precious data source.
I’ll get in touch with some specific project maintainers (e.g. PeerTube or Zappa) but I wanted to post a message here to reach a broader audience. Here are some examples I thought about:
- Recommendation system: classic for videos or whatever
- Privacy-preserving ad system: it is not the most convincing application (even for me) but I saw that there were some discussions about the monetisation so it could be part of the solution.
- ML-based video/audio encoding? I am not knowledgeable about this topic but I read here and there that ML helped to improve the video encoding. We could imagine instances collaborating to train an common ML model that optimises the bit rate for a given quality.
- Federated automatic moderation? I’ve just discover the project Zappa (Announcing the Zappa project) that aims at fighting against misinformation that could rely partially on AI predictions. I don’t know yet how the training is envisioned but I’ll look into it.
Feel free to propose any weird idea, you may have even if you don’t know how to implement it. Note that Federated Learning has an important focus on privacy so the training can be possible even if the data remains of the user devices.
Looking forward to discuss with you!