另有方法是統(tǒng)計(jì)變異程度_第1頁(yè)
另有方法是統(tǒng)計(jì)變異程度_第2頁(yè)
另有方法是統(tǒng)計(jì)變異程度_第3頁(yè)
已閱讀5頁(yè),還剩19頁(yè)未讀, 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說(shuō)明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)

文檔簡(jiǎn)介

1、國(guó)立雲(yún)林科技大學(xué)National Yunlin University of Science and TechnologyThe Evolving TreeAnalysis and ApplicationsAdvisor : Dr. HsuPresenter : Zih-Hui LinAuthor:Jussi Pakkanen, Jukka Iivarinen, and Erkki Oja,Intelligent Database Systems Lab忍 J,展?護(hù)lEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 17, NO. 3, MAY2006Outl

2、ine Motivation Objective ETree ETree-Analysis Experiments ConclusionsWN.Y.U.S.T.IMI.M.Motivation Some of its intrinsic features make it unsuitable for analyzing very large scale problems. This converts complexity control from a global problem to a local one which is simpler.Intelligent Database Syst

3、ems Lab6rI.M.Objective We have analyzed and compared the ETree against many different systems.Intelligent Database Systems Lab6rI.M.ETreeTraining data咖、o,QO/ oX.1)Find the BMU using the search tree2) Update the leaf node locations using the SOM training formulas substituting tree distanee for grid d

4、istanee.w誼 + 1) = w誼)+ g()x(t) - w,f)九?=&exp (_?;*;e/oXMu;/ ;、/3) Increment the BMUs hit counter.4) If the counter reaches the splitting threshold, split the node.Intelligent Database Systems Lab6rI.M.O、/ 1)2)ETree- How to controlling the growthAt the end of each epoch set bit +1) = 7 bi(t) for all

5、nodes z, where 7 日0)1.If the tree has grown less than a specified amount (e.g.,5%) since the last epoch, stop training.oes os ass 0.8 a75 az o.$5 ox 055 as 0.45 m 0塑 as 0.25 0.2 0.15 ai Regularization coefficientaej uoueoESSEo: 1I.M.ETree- Removing LayersRemoved lavers One beneficial feature of most

6、 neural networks is graceful degradation 0.20.10I.M.ETree- better Search for the BMU At every layer we keep the n best subbranches instead of only one.gular BMUTree distances, 2 BMUsTree distancergptimal BML Regular 2 BMUs Optimal BMUI.M. WMhoutPCA Without PCA (optical BMU PCA PCA(

7、opCmaiBMU)Orwal PcaRandomETree- Child Node Initialization The first perturbs the child nodes ponent analysis (PCA).Standard devia nee of subtree sizeIntelligent Database Systems LabI.M.ETree- Optimizing the Leaf Node Locations1) First, we map all training vectors to leaf nodes using the

8、established BMU search.2) Then we move the leaf nodes to the center of mass of their respective data vectors.Large dataset1ag EE EMek mam0.1HomoQonMUS texturebfcntfwrittin digt*11001000900aoo7006005004003002001000EtroeEtree kmaans(b)K-rrieansI.M.Visualization experimentsOlDQQ 3Gau”CIOO&4O-0.0-Q/tCla

9、ssGZIAS&1 cClAS&.0.30 5-0 5ETreeLeaf nodes-o sGOClavu 36 s 1GiaO 5-O.K1Cea* 17Ogar 4SOMKm eans0=50 5-0.6廠小 02) O GPmf-Q-o*050.5-OSOS-0.50.5-05Claool *1datavectorClewi-tq0 5O-O.&-1Q1Class 1OOSooO05CkJt;vtCIUW4C3&6ClassiO;as-OBOS0.5O.S4-SOB06Intelligent Database Systems Lab 厲I.M.Quality of clustering0

10、.5OQuantization error0.45 一0.40.35without the search treeInverse quantization errorDavies-Bouldin indexDBI =Imaxn 翕 i#j徑,(q)+ sg八 k sec)丿Intelligent Database Systems Lab6rWN.Y.U.S.T.IMConclusions ETree5 s performance is quite close to classical, nonhierarchical algorithms but it is noticeably faster, ETree makes implementation and application easier

溫馨提示

  • 1. 本站所有資源如無(wú)特殊說(shuō)明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒(méi)有圖紙預(yù)覽就沒(méi)有圖紙。
  • 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
  • 5. 人人文庫(kù)網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
  • 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請(qǐng)與我們聯(lián)系,我們立即糾正。
  • 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。

評(píng)論

0/150

提交評(píng)論