所需E币: 4
时间: 2019-12-25 15:01
大小: 471.94KB
TraditionalDSPisbasedonalgorithms,changingdatafromoneformtoanotherthroughstep-bystepprocedures.Mostofthesetechniquesalsoneedparameterstooperate.Forexample:recursivefiltersuserecursioncoefficients,featuredetectioncanbeimplementedbycorrelationandthresholds,animagedisplaydependsonthebrightnessandcontrastsettings,etc.Algorithmsdescribewhatistobedone,whileparametersprovideabenchmarktojudgethedata.Theproperselectionofparametersisoftenmoreimportantthanthealgorithmitself.Neuralnetworkstakethisideatotheextremebyusingverysimplealgorithms,butmanyhighlyoptimizedparameters.Thisisarevolutionarydeparturefromthetraditionalmainstaysofscienceandengineering:mathematicallogicandtheorizingfollowedbyexperimentation.Neuralnetworksreplacetheseproblemsolvingstrategieswithtrial&error,pragmaticsolutions,anda"thisworksbetterthanthat"methodology.ThischapterpresentsavarietyofissuesregardingparameterselectioninbothneuralnetworksandmoretraditionalDSPalgorithms.CHAPTERNeuralNetworks(andmore!)26TraditionalDSPisbasedonalgorithms,changingdatafromoneformtoanotherthroughstep-by-stepprocedures.Mostofthesetechniquesalsoneedparameterstooperate.Forexample:recursivefiltersuserecursioncoefficients,featuredetectioncanbeimplementedbycorrelationandthresholds,animagedisplaydependsonthebrightnessandcontrastsettings,etc.Algorithmsdescribewhatistobedone,whileparametersprovideabenchmarktojudgethedata.Theproperselectionofparametersisoftenmoreimportantthanthealgorithmitself.Neuralnetworkstakethisideatotheextremebyusingverysimplealgorithms,butmanyhighlyoptimizedparameters.Thisisarevolutionarydeparturefromthetraditionalmai……