Local Epochs Inefficiency Caused by Device Heterogeneity in Federated Learning

Table 8

The standard deviation of convolutional feature weights ranking.

Features	P100_Conv	V100_Conv	K40_Conv	MeanRankStd
Batchsize	1.7653	1.3045	1.4026	1.4908
Elements_matrix	2.3233	2.0775	1.1578	1.8529
Elements_kernel	1.7904	2.2949	1.4356	1.8403
Channels_in	0.4440	0.4218	0.7599	0.5419
Channels_out	3.0549	2.1915	2.1091	2.4518
Padding	1.6192	1.8036	1.2830	1.5686
Strides	2.4347	2.0012	3.2603	2.5654
Use_bias	1.2580	1.5514	1.3097	1.3730
Opt_SGD	2.3229	3.0940	1.6340	2.3503
Opt_Adadelta	2.2164	2.8144	1.9606	2.3305
Opt_Adagrad	2.7459	2.4789	1.1959	2.1402
Opt_Momentum	2.6090	2.1952	1.1699	1.9914
Opt_Adam	1.4019	2.1550	2.4311	1.9960
Opt_RMSProp	2.0723	2.3218	0.9707	1.7883
Act_relu	1.5234	1.0455	0.9251	1.1647
Act_tanh	1.7225	1.1762	1.4122	1.4370
Act_sigmoid	1.3042	1.1182	1.1107	1.1777