Improving linear separability of classes via feature projection
V. Chandrasekaran, Zhiqiang Liu · 2002
It is well known that a linearly separable set of classes is ideal for a pattern recognition task. The majority of pattern recognition research has been devoted to achieve linear separability of classes by nonlinear input-output mapping. We develop a novel idea of class label separation by projecting each element of the feature vector onto a manifold. The functional characteristics of the manifold associated with each feature type are learnt iteratively from the class label distribution under an optimization criterion. This process attempts to transform an n-dimensional nonlinearly separable feature classification task to an n-dimensional linearly separable problem. The burden of classifying features that are associated with multiple class labels is handled by projections of other discriminating features. This enables fast learning of the classification task by the second stage network which accepts the projected output as its input. If the classification task is modified by an addition of a feature element, the system requires iterative learning of the manifold associated with this new unit only and does not require learning of the whole set of features as seen in conventional neural networks. This iterative knowledge aggregation permits ease of fine tuning and selection of an optimal set of parameters for a given task. The above concept is demonstrated on a set of classification tasks.