This letter analyzes the Fisher kernel from a statistical point of view. The Fisher kernel is a particularly interesting method for constructing a model of the posterior probability that makes intelligent use of unlabeled data (i.e., of the underlying data density). It is important to analyze and ultimately understand the statistical properties of the Fisher kernel. To this end, we first establish sufficient conditions that the constructed posterior model is realizable (i.e., it contains the true distribution). Realizability immediately leads to consistency results. Subsequently, we focus on an asymptotic analysis of the generalization error, which elucidates the learning curves of the Fisher kernel and how unlabeled data contribute to learning. We also point out that the squared or log loss is theoretically preferable-because both yield consistent estimators-to other losses such as the exponential loss, when a linear classifier is used together with the Fisher kernel. Therefore, this letter underlines that the Fisher kernel should be viewed not as a heuristics but as a powerful statistical tool with well-controlled statistical properties.

Recently, Jaakkola and Haussler (1999) proposed a method for constructing kernel functions from probabilistic models. Their so-called Fisher kernel has been combined with discriminative classifiers such as support vector machines and applied successfully in, for example, DNA and protein analysis. Whereas the Fisher kernel is calculated from the marginal log-likelihood, we propose the TOP kernel derived from tangent vectors of posterior log-odds. Furthermore, we develop a theoretical framework on feature extractors from probabilistic models and use it for analyzing the TOP kernel. In experiments, our new discriminative TOP kernel compares favorably to the Fisher kernel.