Local Epochs Inefficiency Caused by Device Heterogeneity in Federated Learning

Table 1

The description of layer features.

Kind of features	Features	Description
Common features	Activation function	The activation function of neuron output, the common ones are sigmoid, tanh, and relu, etc.
	Optimizer	The optimization method of the model, the common ones are SGD, Adadelta, Adagrad, momentum, Adam, and RMS prop, etc.
Dense features	Number of inputs	Since the MLP layers are fully connected, the input of each layer comes from the output of the previous layer.
	Number of neurons	The number of neurons.
Convolutional features	Matrix size	The size of the input data.
	Kernel size	The size of convolutional kernel.
	Input depth	The number of input channels.
	Output depth	The number of output channels.
	Stride size	The convolution step size of convolution kernel.
	Input padding	The number of edge padding after convolution.
Hardware features	GPU clock speed	GPU clock cycle speed.
	GPU memory bandwidth	GPU bandwidth.
	GPU core count	The number of GPU processing units, which represents the number of CUDA cores in NVIDIA GPU.