




版權(quán)說(shuō)明:本文檔由用戶(hù)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、Recent Progress on Active LearningSheng-Jun Huang (黃圣君)Nanjing University of Aeronautics and Astronautics2018-4-22 VALSELearning with Fewer Labeled Data2 years for 4000 sentencesin PennTreebanktime consumingonly experts can provideaccurate annotationshigh expertisebut expensiveLabeled data is import
2、ant Can we learn with fewer labeled data?2Active Learninglabeled dataquery some labelsoracle(annotator)trainmunlabeled dataGoal: train an effective mwith least labeling cost3Active LearningWhich instance to select?Informative instancesRepresentative instancesInformative & representative instances4Re
3、cent ProgressWeak supervisionThe oracle may be noisyor unavailableCost sensitiveCare the cost rather thanthe numberMdependentDifferent ms mayhave diverse needsMore Practical and More Systematic5Recent ProgressWeak supervisionThe oracle may be noisyor unavailableCost sensitiveCare the cost rather tha
4、nthe numberMdependentDifferent ms mayhave diverse needs6Active learning with Weak SupervisionCollaborative labeling from crowdsLabeler quality estimationEnsemble kernel machine classifierRobust to label noisemHua. Collaborative Active Visual Recognition from Crowds A Distributed Ensemble Approach. P
5、AMI 2018.7.Active learning with Weak SupervisionPairwise comparison from noisy labelersLeverage both types of oraclesLower querying complexity under different noise conditionsLabeling oracleComparison oraclemXu. Noise-Tolerant Interactive Learning Using Pairwise Comparisons. NIPS 2017.8Active learni
6、ng with Weak SupervisionSelf-paced active learningSelf-annotation for high-confident instancesOracle annotation for low-confident instancesLin. Active Self-Paced Learning for Cost-Effective and Progressive Face Identification. PAMI 2018.9Active learning with Weak SupervisionActive query from source
7、domainsOracle is not available in the target domainInsufficient labeled data in all domainsOracledomainadaptationSource DomainTarget DomainWang. On Gleaning Knowledge from Multiple Domains for Active Learning. IJCAI 2017.10Unlabeled data Labeled data Unlabeled data Labeled data Recent ProgressWeak s
8、upervisionThe oracle may be noisyor unavailableCost sensitiveCare the cost rather thanthe numberMdependentDifferent ms mayhave diverse needs11Cost-Sensitive Active LearningOracles are cost-sensitiveDifferent oracles have diverse pricesSelecting both instance and oracleAccurate yet cheap annotationsL
9、ow overall qualityLow priceExpert for this queryHigh overall qualityHigh priceLess familiar with itmWho is this ?Huang. Cost-Effective Active Learning from Diverse Labelers. IJCAI 2017.12.Cost-Sensitive Active LearningLabels are cost-sensitiveLabels have hierarchiesBi-objective optimization tobalanc
10、e the cost and informationYan. Cost-Effective Active Learning for Hierarchical Multi-Label Classification. IJCAI 2018.13Cost-Sensitive Active LearningLearning task is cost-sensitiveQuery the cost of predicting a specific labelGuarantee a polynomial improvement onlabel complexity for low noise caseKr
11、ishnamurthy. Active Learning for Cost-Sensitive Classification. ICML 2017.14Recent ProgressWeak supervisionThe oracle may be noisyor unavailableCost sensitiveCare the cost rather thanthe numberMdependentDifferent ms mayhave diverse needs15Active Learning with Deep MsActive madaptationA novel criteri
12、on “distinctiveness”Reuse of pre-trained mLess training datasHuang. Cost-Effective Training of Deeps with Active MAdaptatio. arXiv 2018.16Active Learning with Deep MsActive annotation with deep generative msDeep generative mto create novel instancesOracle directly annotates the decision boundaryHuijser. Active Decision Boundary Annotation with Deep Generative Ms. ICCV 2017.17Active Learning for Various ApplicationsHuman Pose Estimation Liu & Ferrari ICCV17Face Identification Lin. PAMI18Semantic
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(hù)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(hù)上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶(hù)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(hù)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。
最新文檔
- 六一活動(dòng)幼兒園活動(dòng)方案
- 六一活動(dòng)籃球館活動(dòng)方案
- 六一烘焙活動(dòng)方案
- 六一特色活動(dòng)方案
- 六一藝?;顒?dòng)方案
- 六一西瓜活動(dòng)方案
- 六公司美食團(tuán)購(gòu)活動(dòng)方案
- 醫(yī)療護(hù)理考試試題及答案
- 醫(yī)技編制考試試題及答案
- 醫(yī)??荚囋囶}及答案
- JJF 1059.2-2012用蒙特卡洛法評(píng)定測(cè)量不確定度
- GH/T 1093-2014再生資源回收利用網(wǎng)絡(luò)規(guī)范
- 培養(yǎng)學(xué)生核心素養(yǎng)導(dǎo)向下的課堂教學(xué)課件
- 廣西壯族自治區(qū)工程造價(jià)綜合定額答疑匯編2022年11月更新
- 食品供應(yīng)不合格產(chǎn)品退換方案
- 年產(chǎn)12000噸水合肼(100%)項(xiàng)目環(huán)評(píng)報(bào)告書(shū)
- 《有機(jī)波譜分析》期末考試試卷及參考答案
- 最詳細(xì)的整車(chē)開(kāi)發(fā)流程
- 部編版七年級(jí)歷史(下)材料論述題專(zhuān)項(xiàng)訓(xùn)練
- 年產(chǎn)1000噸乳酸的生產(chǎn)工藝設(shè)計(jì)
- 博克服裝CAD制版說(shuō)明操作手冊(cè)(共95頁(yè))
評(píng)論
0/150
提交評(píng)論