12.07.2015 Views

Neural Networks - Algorithms, Applications,and ... - Csbdu.in

Neural Networks - Algorithms, Applications,and ... - Csbdu.in

Neural Networks - Algorithms, Applications,and ... - Csbdu.in

SHOW MORE
SHOW LESS
  • No tags were found...

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

304 Adaptive Resonance TheoryNote the similarity between Eq. (8.19) <strong>and</strong> Eq. (6.13) from Section 6.1.3.Both equations describe a competitive layer with on-center off-surround <strong>in</strong>teractions.In Section 6.1.3, we showed how the choice of the functional form ofg(x) <strong>in</strong>fluenced the evolution of the activities of the units on the layer. 5 Withoutrepeat<strong>in</strong>g the analysis, we assume that the values of the various parameters <strong>in</strong>Eq. (8.19), <strong>and</strong> the functional form of g(x) have been chosen to enhance theactivity of the s<strong>in</strong>gle F 2 node with the largest net-<strong>in</strong>put value from F\ accord<strong>in</strong>gto Eq. (8.16). The activities of all other nodes are suppressed to zero. Theoutput of this w<strong>in</strong>n<strong>in</strong>g node is given a value of one. We can therefore expressthe output values of F 2 nodes asf 1 T, =max{T A ,}VA:uj = f(x 2] ) = { *• \ (8.20)[ 0 otherwiseWe need to clarify one f<strong>in</strong>al po<strong>in</strong>t concern<strong>in</strong>g Figure 8.4. The process<strong>in</strong>gelement <strong>in</strong> that figure appears to violate our st<strong>and</strong>ard of a s<strong>in</strong>gle output per node:The node sends an output of g(x 2 j) to the F 2 units, <strong>and</strong> an output of f(x2j)to the FI units. We can reconcile this discrepancy by allow<strong>in</strong>g Figure 8.4to represent a composite structure. We can arrange for unit Vj to have thes<strong>in</strong>gle output value x 2 j. This output can then be sent to two other process<strong>in</strong>gelements; one that gives an output of g(x2j), <strong>and</strong> one that gives an output off(x 2j ). By assum<strong>in</strong>g the existence of these <strong>in</strong>termediate nodes, or <strong>in</strong>terneurons,we can avoid violat<strong>in</strong>g the s<strong>in</strong>gle-output st<strong>and</strong>ard. The node <strong>in</strong> Figure 8.4 thenrepresents a composite of the Vj nodes <strong>and</strong> the two <strong>in</strong>termediate nodes.Top-Down LTM Traces. The equations that describe the top-down LTM traces(weights on connections from F 2 units to F\ units) should be somewhat familiarfrom the study of Chapter 6:^j = (-ztj + h(xu))f(x 2 j) (8.21)S<strong>in</strong>ce f(x 2 j} is nonzero for only one value of j (one F 2 node, Vj), Eq. (8.21) isnonzero only for connections lead<strong>in</strong>g down from that w<strong>in</strong>n<strong>in</strong>g unit. If the jthF 2 node is active <strong>and</strong> the zth FI node is also active, then Zj, = — z,j + 1 <strong>and</strong> £,;_/asymptotically approaches one. If the jth F 2 node is active <strong>and</strong> the zth F, nodeis not active, then Zjj = —z.-,j <strong>and</strong> z t j decays toward zero. We can summarizethe behavior of z,, as follows:{—Zij + 1 Vj active <strong>and</strong> v, active—Zij Vj active <strong>and</strong> ?;,- <strong>in</strong>active (8.22)0 Vj <strong>and</strong> v\ both <strong>in</strong>activeRecall from Eq. (8.15) that, if F> is active, then v, can be active only ifit is receiv<strong>in</strong>g an <strong>in</strong>put, /,, from below <strong>and</strong> a sufficiently large net <strong>in</strong>put, Vj,5 Caulion! The function t;(.r) <strong>in</strong> this section is the analog of f(j-) <strong>in</strong> Section 6.1.3. In Section 6.1,3,g(.r) was used to mean j-~'/(.r). In this section,

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!