GENERALIZATION OF LINEAR REGRESSION PROBLEMS - Neural-based Orthogonal Data Fitting: The EXIN Neural Networks - page 199

1.5

x 1

1

0.5

x 2

0

50 100 150 200 250

α − >

Figure 5.27 Weighted TLS method: plot of x 1 and x 2 vs. α .

Table 5.3 Total computational cost (flops) of the inverse iteration and scheduling
methods for the benchmark problem

m

50 100 250 400 1000

Inverse iteration 31,123 122,596 756,316 1,928,876 12,022,016

Scheduling 46,175 313,799 451,397 632,009 632,747

Table 5.4 Overdetermined linear system benchmark a

ζ α 0 γ Time (ms)

GeTLS EXIN seq. 0 0.25 0.45 0.4

GeTLS EXIN seq. hyp. 0.035 0.002 2

GeTLS EXIN batch hyp. 0.0046 0 194

GeTLS EXIN batch 0 0.4 0.6845 4.2

GeTLS EXIN batch 0.5 0.03 0.0038 10

SCG GeTLS EXIN 0 — — 0.1

SCG GeTLS EXIN 0.5 — — 0.8

SCG GeTLS EXIN 0.99 — — 1.6

BFGS GeTLS EXIN 0 — — 0.1

BFGS GeTLS EXIN 0.5 — — 0.1

BFGS GeTLS EXIN 0.99 — — 0.1

a seq., sequential; hyp., hyperbolic scheduling.