MCA EXIN NEURON - Neural-based Orthogonal Data Fitting: The EXIN Neural Networks - page 81

MCA EXIN

LUO

OJAn

OJA

OJA

MCA EXIN

iterations

Figure 2.17 Line fitting for noise variance equal to 0.5: zoom of the plot of the weights for
the transient analysis. ( See insert for color representation of the figure .)

Table 2.3 Total cost of the MCA learning laws for autocorrelation matrices
of increasing size ∗

κ 2

dim EXIN LUO FENG OJA

3 20 42,726 37,466 41,133 20,888

5 40 65,966 59,131 60,345 42,942

7 60 183,033 640,074 742,346 div.

8 70 205,968 706,075 830,000 div.

∗

9 80 228,681 965,544 div.

10 90 252,002 1,061,098 div. div.

15 140 366,255 div. div. div.

25 240 2,003,755 div. div. div.

50 490 3,725,868 div. div. div.

75 740 4,488,496 div. div. div.

100 990 6,258,001 div. div. div.

∗ Inaccurate result; div., divergence.

divergence). Figure 2.21 shows the good results obtained by MCA EXIN (here
λ n = 0 . 01).
To avoid starting with too high initial conditions, a second group of exper-
iments has been done where the initial weight vector is the same as in the
first experiment, but divided by 10 times its norm. Table 2.4 shows the results
about divergence of all MCA laws except MCA EXIN, which always converges.
Obviously, the divergence happens later than in the first group of experiments
because the weights are lower. Figures 2.22 and 2.23 show some results. A
possible interpretation of the divergence in these experiments is the following: