Abstract
The Fisher kernel, which refers to the inner product in the feature space of the Fisher score, has been known to be a successful tool for feature extraction using a probabilistic model.
If an appropriate probabilistic model for given data is known, the Fisher kernel provides a discriminative classifier with good generalization.
However, if the distribution is unknown, it is difficult to obtain an appropriate Fisher kernel.
In this paper, we propose a new nonparametric Fisher-like kernel derived from fuzzy clustering instead of a probabilistic model,
noting that fuzzy clustering methods such as a family of fuzzy c-means are highly related to probabilistic models.
Numerical examples show the effectiveness of the proposed method.
