keyword

Fisher kernel framework

The Fisher kernel framework is a machine learning method that bridges generative probabilistic models and discriminative classifiers by deriving a similarity measure from the parameters of an underlying generative distribution. It characterizes an arbitrary data sample by computing the gradient of its log-likelihood with respect to the generative model parameters, which is typically scaled by the Fisher information matrix. This process transforms variable-sized or complex structured inputs into a fixed-dimensional gradient representation, often called a Fisher vector, that reflects how the observed data deviates from the baseline generative model, such as a Gaussian mixture model. In pattern recognition and computer vision, this framework allows unordered sets of local descriptors to be aggregated into rich, global representations that can be directly analyzed, indexed, or classified using standard discriminative algorithms.

1 item