Glossary

Here below the glossary used for Fed-BioMed :

• experiment : orchestrates the rounds during the federated learning, on the available nodes

  • it includes : training plan, model, federated trainer, training parameters, model parameters, set of input data, results
  • an experiment is unique (cannot be replayed) and is over when converged
  • status : running and then done

• training : as commonly used in ML, process of feeding a model with data to improve its accuracy on some task.

• validation : process of giving a heuristic information on the accuracy of a model during training.

• testing : process of assessing the accuracy of a model after training, on holdout samples different from the one that were used for training. Not implemented yet in Fed-BioMed.

• job : a pure researcher notion. Interface between the researcher and the nodes of an experiment for one single request (short lived, compared to experiment). It triggers the local work for all sampled nodes for one round.

• round : everything included in choice of the nodes, perform local work on the nodes, sending back whatever information is required, server performs the aggregation
  • current Round() class on node corresponds to local work
• parameter update : an update of the ML model parameters during the training loop, which usually corresponds to the processing of one batch of data

• epoch : a number of parameter updates equivalent to processing the entire dataset exactly once

• researcher (technical) : entity that defines and executes an experiment

• node (technical) : entity with tagged datasets that replies to researcher queries and performs local work