Data compression without loss, statistics and adaptive.
Principle
The weighting of context uses several modelings of context to evaluate the probability of the various symbols.
By gathering these various modelings having each clean advantage, the weighting of context is judicious to offer a higher reliability in the prediction than each taken modeling separately.
The various entries of a CM are in general adapted to different types of data. For example a CM can mix the exits of a predictor specialized for the text and of another specialized for the images, each one being with priori more powerful on the type of data for which it is specialized; it should thus approach the performances of the predictor specialized for the text on the text and those of the predictor specialized for the images on the images.
Properties
The weighting of contexts is a symmetrical algorithm. That means that it makes the same thing with compression and decompression. That also means that its speed is the same one in both cases (if one does not take account of subtleties of the input-outputs), and that the quantity of memory necessary (to store the mixeurs, the predictors and their contexts) is identical.
Performances
The compression ratios obtained by CMs among the best are obtained today and this, whatever the type of data to be compressed (except, of course random data).
The quantity of memory necessary is in general important. Each predictor has his own needs and the mixeurs can also require some resources (for example in the case of a network of neurons).
Speed is the weak point of CMs. They are all the more slow as the number of predictors (or contexts) is important. Certain mixeurs " intelligents" decontaminate dynamically the nonrelevant predictors and contexts to accelerate calculations.
A CM is easily parallélisable because, in practice, the predictors are almost always independent.
