欢迎来到天天文库
浏览记录
ID:62066927
大小:2.32 MB
页数:138页
时间:2021-04-14
《最新Chapter 3 Mapping Networks - 醫學影像處理實驗室MIPL-药学医学精品资料.ppt》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、最新Chapter3MappingNetworks-醫學影像處理實驗室MIPL-药学医学精品资料IntroductionMappingNetworksAssociativememorynetworksFeed-forwardmultilayerperceptronsCounter-propagationnetworksRadialBasisFunctionnetworks.MappingANN(.)xyAssociativeMemoryNetworks記憶容量(memorycapability)對於資訊系統而言,記憶(remember)並且降低(reduce)所儲存的資訊量是很重
2、要的。儲存的資訊必須被適當的存放在網路的記憶體中,也就是說,給予一個關鍵的(key)輸入或刺激(stimulus),將從聯想式記憶體(associativememory)中擷取記憶化的樣本(memorizedpattern),以適當的方式輸出。GeneralLinearDistributedAssociativeMemory(cont.)Input(keyinputpattern)Output(memorizedpattern)對一個n-dimension的神經架構,可聯想(associate)h個pattern。h<=n。實際上是h3、arDistributedAssociativeMemory(cont.)Keyvectorxk和記憶向量yk間的線性對應可寫成:其中W(k)為權重矩陣(weightmatrix)從這些權重矩陣,可建構MemorymatrixMdescribesthesumoftheweightmatricesforeveryinput/outputpair.Thiscanbewrittenasthememorymatrixcanbethoughtofasrepresentingthecollectiveexperience.(3.3)(3.4)(3.5)CorrelationMat4、rixMemoryEstimationofthememorymatrixOuterproductruleEachouterproductmatrixisanestimateoftheweightmatrixW(k)thatmapstheoutputpatternykontotheinputpatternxk.Anassociativememorydesignedinouterproductruleiscalledacorrelationmatrixmemory(3.6)(3.7)CorrelationMatrixMemory(cont.)當某keypattern輸入時,希望memo5、rymatrix能recall適當的memorizedpattern,假設estimatedmemorymatrix已經由(3.6)學習了h次,則輸入第q個keyinputxq,欲求memorizedyq(3.8)式可改寫成(將(3.6)代入(3.8))假設每個keyinputvector具有normalizedunitlength.(3.8)(3.9)(3.10)CorrelationMatrixMemory(cont.)此normalizedunitlength可藉由除以Euclideannorm來求得:因此,(3.9)可改寫成其中(3.11)(3.12)6、(3.13)Noise,crosstalkDesiredresponseCorrelationMatrixMemory(cont.)從(3.13)當網路引進associatedkeypattern(y=yq,zq=0)若y=yq表示輸入xq可精確的得到其對應的輸出yq。若zq不為0,表示有noise或crosstalk。若keyinput之間為orthogonal,則crosstalk為0因此可得到perfectmemorizedpattern。若輸入的keyinput之間存在linearlyindependent,則可在計算memorymatrix之前先使用Gram-Schmid7、torthogonalization來產生orthogonalinputkeyvectors。(3.14)CorrelationMatrixMemory(cont.)ThestoragecapacityoftheassociativememoryisThestoragelimitdependsontherankofthememorymatrix.CorrelationMatrixMemory(cont.)Example3.1Autoassociativememory
3、arDistributedAssociativeMemory(cont.)Keyvectorxk和記憶向量yk間的線性對應可寫成:其中W(k)為權重矩陣(weightmatrix)從這些權重矩陣,可建構MemorymatrixMdescribesthesumoftheweightmatricesforeveryinput/outputpair.Thiscanbewrittenasthememorymatrixcanbethoughtofasrepresentingthecollectiveexperience.(3.3)(3.4)(3.5)CorrelationMat
4、rixMemoryEstimationofthememorymatrixOuterproductruleEachouterproductmatrixisanestimateoftheweightmatrixW(k)thatmapstheoutputpatternykontotheinputpatternxk.Anassociativememorydesignedinouterproductruleiscalledacorrelationmatrixmemory(3.6)(3.7)CorrelationMatrixMemory(cont.)當某keypattern輸入時,希望memo
5、rymatrix能recall適當的memorizedpattern,假設estimatedmemorymatrix已經由(3.6)學習了h次,則輸入第q個keyinputxq,欲求memorizedyq(3.8)式可改寫成(將(3.6)代入(3.8))假設每個keyinputvector具有normalizedunitlength.(3.8)(3.9)(3.10)CorrelationMatrixMemory(cont.)此normalizedunitlength可藉由除以Euclideannorm來求得:因此,(3.9)可改寫成其中(3.11)(3.12)
6、(3.13)Noise,crosstalkDesiredresponseCorrelationMatrixMemory(cont.)從(3.13)當網路引進associatedkeypattern(y=yq,zq=0)若y=yq表示輸入xq可精確的得到其對應的輸出yq。若zq不為0,表示有noise或crosstalk。若keyinput之間為orthogonal,則crosstalk為0因此可得到perfectmemorizedpattern。若輸入的keyinput之間存在linearlyindependent,則可在計算memorymatrix之前先使用Gram-Schmid
7、torthogonalization來產生orthogonalinputkeyvectors。(3.14)CorrelationMatrixMemory(cont.)ThestoragecapacityoftheassociativememoryisThestoragelimitdependsontherankofthememorymatrix.CorrelationMatrixMemory(cont.)Example3.1Autoassociativememory
此文档下载收益归作者所有