Self-Organizing Maps for Anatomical Joint Constraint Modelling

Anatomical Joint Constraint Modelling with Rigid Map Neural Networks

“…An increase in the range (angle between the virtual limb and the z-axis) of the constrained region results in a decrease in performance, as shown in Figure 1. The resulting corrections, however, are inferior to those of the SOM (from our earlier work [19]) using the same training data, training iterations and a similar number of output nodes (625) as shown in Figure 1. Increasing the number of training epochs produced an increases in performance, which attenuates rapidly as the number of epochs increases.…”

“…Previous results with the SOM [19] network showing improved results with an increase in the data set size are not echoed in the results for the Rigid Map, suggesting that the other factors (possibly the limited output node density) limit further improvements in performance.…”

“…To overcome this issue, ANNs trained using unsupervised techniques such as competitive learning have been proposed. SOMs have been trained using competitive learning to implicitly model joint constraints using only valid orientations expressed as unit quaternions [19,20]. The weights of the output nodes are trained via competative learning to represent the training data while preserving the topography of the input space.…”

“…The network responds to a given input orientation with the closest orientation in its model of the input data. This can be used directly for correction [19] or as a target for an iterative approach [20].…”

Evolved Topology Generalized Multi-layer Perceptron (GMLP) for Anatomical Joint Constraint Modelling

“…Research has shown that an SVM classifier can be used to separate valid orientations from those requiring correction [29]. Recent research [30] has looked employ unsupervised learning more applicable to use with recorded actor or patient specific data (such as that gathered by Herda, Urtasun and Fua [26]). …”