Data compression with controllable semantic loss

Abstract

This invention describes apparatuses/methods for data compression/decompression, wherein input observation data is classified into types of semantic objects according to one or more criteria and a compression technique is applied to the object types based on their ease of reconstruction via a compression model to obtain compressed data. The classification may be performed based on the performance of the compression model. A suitable descriptive prompt may be generated, e.g., based on a textual inversion technique, for a data object belonging to a data object type that can be well reconstructed. Furthermore, the invention may be applied for personalized data compression/construction/decompression/reconstruction, wherein an encoder/decoder takes a suitable personalized descriptive prompt (or program) and/or a personalized reconstruction model and/or a personalized policy to generate personalized content according to the user's preferences and/or profile.

Dependencies are inferred from the claim text.