Nowadays the number of data stream applications and domains, where dynamism and speed truly matter, is increasing. In practice those streams represent dynamic data flows, coming from different sources, where their content evolves in time. Moreover, the complexity of current digital applications, and those of the near future, is constantly increasing due to a combination of aspects such as the large number of sources, the non-linearity of certain processes, the distribution of knowledge and control, the time response, the strong dynamics of its environment, the unpredictability of interactions, etc. The aforementioned complexity opens new research challenges about the generation and processing (learning) of these streams, especially in distributed, heterogeneous and collaborative environments. This claims for distributed solutions capable of processing and learning from data in a decentralized way as, for instance, federated learning. Furthermore, distributed learning brings several advantages as reducing the computational complexity (since it is shared among different nodes), minimizing the load and cost of communications (transmission of parameters instead of the full data), or increasing the security (knowledge is shared among different machines).
Based on the previous motivations, the goal of this internship is to implement a decentralized federated algorithm applied to distributed streaming contexts. For this, the student will rely on STREAMER. Developed in LI3A laboratory (CEA List) premises, STREAMER is a cutting-edge data stream processing (Complex Event Processing) framework devoted to analyzing sequential data for electrical or industrial systems. For the completion of this internship, the student will work on three main tasks:

1. Adapt STREAMER to work in a distributed fashion (communications need to be implemented, kernel preparation, etc.).
2. Propose and implement in STREAMER a distributed and decentralized algorithm (based on federated learning) that also deals with streaming challenges (incremental learning, no data storage, single pass, concept drift…).
3. Start exploring the ways of making such kind of algorithms robust to unexpected difficulties such as noise in the data, sources that stop the communication, corrupted data, etc.

 
This internship may be extended to a 3 years PhD.

Keywords: Data Stream, Machine Learning, Federated Learning.

References:
[1] https://www.influxdata.com/, https://kafka.apache.org/, https://redis.io/

[2] Muthukrishnan, S. (2005). Data streams: Algorithms and applications. Foundations and Trends® in Theoretical Computer Science, 1(2), 117-236.
[3] Palpanas, Themis. "Data series management: The road to big sequence analytics." ACM SIGMOD Record 44.2 (2015): 47-52.
[4] Google AI blog: https://ai.googleblog.com/2017/04/federated-learning-collaborative.html