Collaborative Learning

In artificial intelligence, collaborative learning is the paradigm in which several learning entities—models, devices, agents or nodes—train jointly and in coordination, sharing knowledge (parameters, gradients or representations) rather than working in isolation. A clarification up front: the same term has a pedagogical sense—people learning in groups—that is not the subject of this entry.

An umbrella, with federated learning at its center

Collaborative learning is not a single algorithm, but an umbrella of distributed approaches. Its most prominent case—so much so that it is often used as a synonym—is federated learning: training a shared model without centralizing the data, so that each client (a phone, a hospital, a bank) trains locally with its private data and sends only the model updates to the server, which aggregates them into a global model.

Concepts it relates to

It is related to distributed learning, broader, which splits training across several nodes—federated learning being a case with decentralized, heterogeneous data—and to multi-agent learning. It should not be confused with collaborative filtering: despite the name, that is a recommender-system technique that predicts a user's tastes from those of many others, not a joint training of several models.

Benefits and challenges

Its advantages are exploiting data and resources scattered in silos and protecting privacy, since the raw data do not leave their origin. Its challenges are the cost of communication between nodes, the heterogeneity of the data and devices, and governance: who validates the updates and how the global model is defended against possible attacks.