LearningDeepArchitecturesforAI

Theoretical results suggest that deep architectures, composed of multiple levels of non-linear operations, may be necessary to learn complicated functions that can represent high-level abstractions in fields such as vision, language, and other AI-level tasks. These architectures can be seen in neural nets with many hidden layers or in complicated propositional formulae, indicating the importance of depth in learning such complex tasks.

作者简介

Theoreticalresultssuggestthatinordertolearnthekindofcomplicatedfunctionsthatcanrepresenthigh-levelabstractions(e.g.,invision,language,andotherAI-leveltasks),onemayneeddeeparchitectures.Deeparchitecturesarecomposedofmultiplelevelsofnon-linearoperations,suchasinneuralnetswithmanyhiddenlayersorincomplicatedpropositionalformulaere-usingmanysub-formulae.Searchingtheparameterspaceofdeeparchitecturesisadifficulttask,butlearningalgorithmssuchasthoseforDeepBeliefNetworkshaverecentlybeenproposedtotacklethisproblemwithnotablesuccess,beatingthestateof-the-artincertainareas.Thismonographdiscussesthemotivationsandprinciplesregardinglearningalgorithmsfordeeparchitectures,inparticularthoseexploitingasbuildingblocksunsupervisedlearningofsingle-layermodelssuchasRestrictedBoltzmannMachines,usedtoconstructdeepermodelssuchasDeepBeliefNetworks.