2022–, with F. Giannini, A. Cossu and D. Bacciu.
Streaming Continual Learning is a unifying paradigm that combines Streaming Machine Learning, Continual Learning and Time Series Analysis to learn from non-stationary data streams with temporal dependence — adapting rapidly to change without forgetting previously acquired knowledge.
The two communities it draws on each solve half of the problem. Streaming Machine Learning adapts quickly after a concept drift, but discards what it knew before. Continual Learning preserves earlier knowledge, but assumes tasks arrive in tidy, labelled batches. Real problems need both abilities at once, on data that is neither tidy nor independent across time. This became a distinct research direction with Federico Giannini’s PhD and has since grown into a small community of its own: the 1st Streaming Continual Learning Bridge at AAAI 2026 brought the two fields together for the first time, with CapyMOA and Avalanche as its hands-on frameworks.
Selected works
-
A practical guide to streaming continual learning
Andrea Cossu, Federico Giannini, Giacomo Ziffer, Alessio Bernardo, Alexander Gepperth, Emanuele Della Valle, Barbara Hammer, Davide Bacciu
Neurocomputing, 2026Defines Streaming Continual Learning as the meeting point of two communities: Streaming Machine Learning, which adapts quickly after a concept drift, and Continual Learning, which preserves what was learned before. The paper argues that real-world problems need both abilities and shows experimentally that neither field alone achieves them.
-
Proceedings of the 1st Streaming Continual Learning Bridge at AAAI (StreamingCL 2026)
Heitor Murilo Gomes, Andrea Cossu, Federico Giannini, Anton Lee, Nuwan Gunasekara, Emanuele Della Valle (eds.)
CEUR-WS.org, 2026Proceedings of the 1st Streaming Continual Learning Bridge at AAAI 2026, which brought the Streaming Machine Learning and Continual Learning communities together; I co-organized the event and co-edited the volume.
-
Dynamic continuous progressive neural networks for evolving streaming time series
Federico Giannini, Giacomo Ziffer, Emanuele Della Valle
Data Min. Knowl. Discov., 2026Journal extension of cPNN that makes the progressive architecture dynamic, so that the network can keep learning from evolving time series streams without growing at every drift.
-
Don’t Look Back in Anger: MAGIC Net for Streaming Continual Learning with Temporal Dependence
Federico Giannini, Sandro D’andrea, Emanuele Della Valle
IEEE BigData 2025, Macau, 2025Proposes MAGIC Net, which, after each drift, decides online whether to reuse the existing recurrent network through learned masks or to grow it, adapting faster than cPNN with a much smaller memory footprint.
-
Don’t drift away: Advances and Applications of Streaming and Continual Learning
Andrea Cossu, Davide Bacciu, Alessio Bernardo, Emanuele Della Valle, Alexander Gepperth, Federico Giannini, Barbara Hammer, Giacomo Ziffer
ESANN 2025, Bruges, 2025Introductory paper of the ESANN 2025 special session on Streaming Continual Learning: an overview of recent advances and applications that bring together streaming and continual learning.
-
cPB: Continuous Piggyback for Streaming Continual Learning with Temporal Dependence
Reza Paki, Federico Giannini, Emanuele Della Valle
ECML PKDD 2025 Workshops, Porto, 2025cPB brings the Piggyback Continual Learning strategy to continuous recurrent networks: learnable masks over a pretrained network let it reuse previous knowledge for each new concept without forgetting, offering a compact alternative to cPNN.
-
MAcPNN: Mutual Assisted Learning on Data Streams with Temporal Dependence
Federico Giannini, Emanuele Della Valle
IEEE BigData 2024, Washington DC, 2024Proposes Mutual Assisted Learning, inspired by Vygotsky’s sociocultural theory: edge devices, each running a cPNN, learn autonomously and ask peers for help only when needed, cutting communication compared with federated learning.
-
cPNN: Continuous Progressive Neural Networks for Evolving Streaming Time Series
Federico Giannini, Giacomo Ziffer, Emanuele Della Valle
PAKDD 2023, Osaka, 2023Introduces continuous Progressive Neural Networks (cPNN), recurrent networks that handle concept drift, temporal dependence, and catastrophic forgetting together by adding a new column for each concept and reusing past knowledge to learn the new one faster.