基于结构化学习的数据压缩方法分析.pptx

Structured Modeling and Learning in Generalized Data Compression and Processing;Sparse Representation;Multimedia Communication; ;Heterogeneous Data Compression: Data;Heterogeneous Data Compression: Framework;Context is suffix of data symbols in classical context modeling. Define the set of subsequences whose suffix is s. Disjoint property: no string in the context set is a suffix of any other string in this set, or for , ; Exhaustive property: each subsequence of data symbols can find its suffix in the context set, or . ; Graphical probabilistic model Structured prediction model can be represented in the form of graph. Each node for the random variable to predict and each edge for the interdependencies among nodes. Estimating joint or conditional distribution for its nodes. Learning methods Markov random field (MRF) Max-margin Markov network: MRF + SVM Reasoning algorithms Belief propagation (BP) Expectation propagation (EP);Structured Probabilistic Model: Motivation;10;Intuition 2: Complex structure for heterogeneous data;Contribution： Structure consistence → Structured Set;Structured Probabilistic Model: Example;Theoretical support: Sequential Source Coding;Contribution;Contribution; Heterogeneous data Genomic data Long repeats with exception of insertion, deletion, and substitution. Image and Video Spatially spanned along the structures, e.g. edge, texture, and etc. Executables Interlaced data streams, e.g. opcodes, immediate data, and etc.;Definition: Heterogeneous Data;Clue ;;?;;Definition: Context Set;Modeling Prediction: Model Graph;Generalized Context Modeling Model Tree example;Representation Prediction: Model Graph;Model tree to represent generalized context models and their successive relationship. Minimum description length (MDL) principle to select optimal class of context models. Normalized maximum likelihood (NML) to determine optimal weighted probab