Context-Dependent Thinning 27
Figure 7: (A) Overlap of thinned composite codevectors and (B) Overlap of component subsets in
thinned composite codevectors - versus the overlap of the corresponding unthinned composite
codevectors for various versions of thinning procedures. CDTadd - the additive, CDTsub - the
subtractive, CDTadd-sl - the self-exclusive additive, CDTsub-sl - the self-exclusive subtractive CDT
procedure, Paired-M - permutive conjunctive thinning; the densities of component codevectors are
chosen to obtain M of 1s in the thinned codevector. For component codevectors, N=100000, M≈1000.
The number of component codevectors: 5. The results are averaged over 50 runs with different random
codevectors.
Figure 8: (A) Overlap of thinned composite codevectors and (B) Overlap of component subsets in
thinned composite codevectors for various "depth" of additive (CDTadd) and subtractive (CDTsub)
procedures of Context-Dependent Thinning. There are five components in the composite item. Therefore
the input composite codevector includes approximately 5M of 1s. The composite codevector is thinned
to have from 4M to M/4 of 1s. Two curves for thinning depth M are consistent with the corresponding
curves in Figure 7. For all component codevectors, N=100000, M≈1000. The results are averaged over
50 runs with different random codevectors.