The relational approach to dependency estimation entails the selection of a sufficiently compact 'relevance' subset of training-set objects with which any newly occurring object may be compared in order to estimate its hidden target characteristics. If several comparison modalities are available, a 'relevance' subset of these may additionally have to be chosen via an appropriate selection criterion. Typically, the level of selectivity will constitute a free parameter, and in traditional approaches, multiple training repetitions would be required to determine this value via cross-validation. To avoid this, we seek to algorithmically emulate the cross-validation process using conservative assumptions as to the nature of the unknown probability distribution that produced the training set. We term this approach 'non-enumerative cross-validation', and demonstrate that the classical Akaike Information Criterion is a specific case of it under naive assumptions. The application of this non-enumerative cross-validation strategy is demonstrated on the standard multikernel data set, "chicken-pieces", treated from the perspective of relational discriminant analysis.

In multimodal biometric information fusion, it is common to encounter missing modalities in which matching cannot be performed. As a result, at the match score level, this implies that scores will be missing. We address the multimodal fusion problem involving missing modalities (scores) using support vector machines (SVMs) with the neutral point substitution (NPS) method. The approach starts by processing each modality using a kernel. When a modality is missing, at the kernel level, the missing modality is substituted by one that is unbiased with regards to the classification, called a neutral point. Critically, unlike conventional missing-data substitution methods, explicit calculation of neutral points may be omitted by virtue of their implicit incorporation within the SVM training framework. Experiments based on the publicly available Biosecure DS2 multimodal (scores) data set show that the SVM-NPS approach achieves very good generalization performance compared to the sum rule fusion, especially with severe missing modalities.

Multi-modal pattern recognition must frequently truncate the set of initially available modalities. When a kernel-based approach is adopted within each modality, the problem of modality selection becomes mathematically analogous to that of wrapper-based feature selection. In this paper, we revise two implicitly wrapper based methods of SVM-embedded selective kernel combination, the Relevance and Support Kernel Machines, so as to equip them with the ability to preset the desired level of feature-selectivity. Hence, a continuous axis of nested feature selection models is obtained, ranging from the absence of selectivity to the selection of single features. We thus unite the distinct processes of selection and classification within the two techniques in manner suitable for general application within Kernel-based multi-modal pattern recognition.