Matches in SemOpenAlex for { <https://semopenalex.org/work/W3205251036> ?p ?o ?g. }
Showing items 1 to 51 of
51
with 100 items per page.
- W3205251036 abstract "Message-passing algorithms consist of a parallelised computing scheme to estimate the marginals of a high-dimensional probability distribution. They have been used in various areas where the statistics of a large number of interacting variables have to be studied, including statistical physics, artificial intelligence, decoding in information theory. This thesis desctibes the algebraic and topological structures in which message-passing algorithms naturally take place. In most applications, the probability distribution p is defined by a Markov field or graphical model, i.e. as a product of local factors depending only on small subsets of interacting variables. Equivalently, the total energy H = - ln p (or log-likelihood) is a sum of local interaction potentials. Embedding potentials as the 0-degree subspace of a chain complex of local observables, we show that this constraint is equivalent to assuming a one-to-one parametrisation of total energies by homology classes of interaction potentials. The 1-chains of this complex are statistical analogs of heat fluxes. The evolution of potentials up to the boundary of a heat flux, preserving total energy, produces new belief propagation algorithms as continuous-time diffusion equations. The consistency of pseudo-marginals one seeks to approximate is reciprocally translated by a cohomological constraint, demanding that a collection of local probability distributions has a vanishing differential. The local potentials and local distributions form a pair of dual variables put in correspondence by a smooth non-linear functional, essentially composing the exponential with a sum over subsets, called zeta transform in combinatorics. Pairs at the intersection of the energy conservation and marginal consistency constraint surfaces are shown to be in correspondence with both the stationary points of belief propagation algorithms and the critical points of a local approximation of free energy, computed by the Bethe-Kikuchi combinatorial method, a truncation of the Mobius inversion of the zeta transform." @default.
- W3205251036 created "2021-10-25" @default.
- W3205251036 creator A5049412905 @default.
- W3205251036 date "2020-12-14" @default.
- W3205251036 modified "2023-09-26" @default.
- W3205251036 title "Message-passing algorithms and homology : from thermodynamics to statistical learning" @default.
- W3205251036 hasPublicationYear "2020" @default.
- W3205251036 type Work @default.
- W3205251036 sameAs 3205251036 @default.
- W3205251036 citedByCount "0" @default.
- W3205251036 crossrefType "dissertation" @default.
- W3205251036 hasAuthorship W3205251036A5049412905 @default.
- W3205251036 hasConcept C105795698 @default.
- W3205251036 hasConcept C114614502 @default.
- W3205251036 hasConcept C118615104 @default.
- W3205251036 hasConcept C149441793 @default.
- W3205251036 hasConcept C33923547 @default.
- W3205251036 hasConcept C98763669 @default.
- W3205251036 hasConceptScore W3205251036C105795698 @default.
- W3205251036 hasConceptScore W3205251036C114614502 @default.
- W3205251036 hasConceptScore W3205251036C118615104 @default.
- W3205251036 hasConceptScore W3205251036C149441793 @default.
- W3205251036 hasConceptScore W3205251036C33923547 @default.
- W3205251036 hasConceptScore W3205251036C98763669 @default.
- W3205251036 hasLocation W32052510361 @default.
- W3205251036 hasOpenAccess W3205251036 @default.
- W3205251036 hasPrimaryLocation W32052510361 @default.
- W3205251036 hasRelatedWork W2008883982 @default.
- W3205251036 hasRelatedWork W2120894224 @default.
- W3205251036 hasRelatedWork W2130550844 @default.
- W3205251036 hasRelatedWork W2949783835 @default.
- W3205251036 hasRelatedWork W2950219171 @default.
- W3205251036 hasRelatedWork W2950582841 @default.
- W3205251036 hasRelatedWork W2963534231 @default.
- W3205251036 hasRelatedWork W2963795759 @default.
- W3205251036 hasRelatedWork W2964024653 @default.
- W3205251036 hasRelatedWork W2984050930 @default.
- W3205251036 hasRelatedWork W3043621011 @default.
- W3205251036 hasRelatedWork W3098128495 @default.
- W3205251036 hasRelatedWork W3102094528 @default.
- W3205251036 hasRelatedWork W3103073721 @default.
- W3205251036 hasRelatedWork W3103161841 @default.
- W3205251036 hasRelatedWork W3103675751 @default.
- W3205251036 hasRelatedWork W3105177482 @default.
- W3205251036 hasRelatedWork W3126279954 @default.
- W3205251036 hasRelatedWork W3176098767 @default.
- W3205251036 hasRelatedWork W2740833984 @default.
- W3205251036 isParatext "false" @default.
- W3205251036 isRetracted "false" @default.
- W3205251036 magId "3205251036" @default.
- W3205251036 workType "dissertation" @default.