Unsupervised Learning of Overlapping Image Components Using Divisive Input Modulation

Table 3

Details of the training procedure used for each of the benchmarking algorithms tested. In all cases the parameters values listed were those found to produce the best results. Parameter values were kept constant across variations in the task. All algorithms except 𝛽=0.01 use an online learning procedure and hence training time is measured in cycles, whereas for 𝜈=0.1 training time is measured in epochs. See the caption of Table 2 for further details.

Algorithm	Training time	Iterations	Weight initialisation	Parameter values
𝚑𝚊𝚛𝚙𝚞𝚛𝟸	200 000 cycles	n/a	mean = 132, std = 18	𝛽=0.0025, 𝜇=0.05
𝜃=0.5	20 000 cycles	50	mean = 𝚗𝚖𝚏𝚜𝚌, std = 12	18, 𝑠𝑊=0.5, 𝑠𝑌=
𝚍𝚒	2 000 epochs	n/a	mean = 136, std = 0.001	𝛽^+=0.25, 𝛽^−=0.25 none
𝚍𝚒𝚖	20 000 cycles	25	mean = 116, std = 164	𝛽=0.05, 𝑠=2
𝑠=3	20 000 cycles	50	mean = 𝑠=4, std = 𝚍𝚒𝚖	#1