FedFF

Investigation of the performance of the Forward-Forward algorithm in a federated learning context.

Simulation parameters

Client model parameters:

• LAYER_EPOCHS:   training epochs for a single layer
• BIAS_THRESHOLD:   model hyper-parameter
• LEARN_RATE:   initial learning rate for ADAM optimizer
• WEIGHT_DECAY:   weight decay value
• MODEL_UNITS:   list containing the number of the units of each layer

Federated algorithm parameters:

• MODEL_EPOCHS:   training epochs for the client model
• NUM_ROUNDS:   number of communication rounds
• NUM_CLIENTS:   number of clients in the network
• C_RATE:   fraction of selected clients taking part to each round

Dataset processing parameters:

• NUM_REPEAT:   training dataset repetitions
• BATCH_SIZE:   local minibatch size for client update
• SHUFFLE_BUF:   dataset shuffle buffer size to sort samples
• PREFETCH_BUF:   size of prefetch buffer