Biometric Authentication: A Machine Learning Approach

10.3. Hierarchical Neural Networks for Sensor Fusion

Several categories of fusion layers have been studied by Abidi and Gonzalez [1]. In the present context, the classes-in-expert hierarchical NN described in Section 6.4.2 can be naturally extended to cover the sensor-fusion network. To this end, the definition of experts needs to be properly expanded. Local experts now include not only the adaptive-trained type but also the predetermined type. For the former, the feature space (often a certain local region) represented by an expert is adaptively trained in an a posteriori fashion. The model parameters depend very much on the initial condition to which the local expert is assigned. As for the predetermined experts, each local expert has fixed ...