东南大学 张敏灵 - 弱监督机器学习范式

吉悦心

2017/12/18 发布于 技术 分类

张敏灵,东南大学计算机科学与工程学院教授。分别在2001年、2004年和2007年于南京大学计算机科学与技术系获学士、硕士和博士学位。主要研究领域为机器学习、数据挖掘。现任中国计算机学会人工智能与模式识别专委会常务委员、中国人工智能学会机器学习专委会秘书长等。担任《ACM Trans. IST》、《Frontiers of Computer Science》、《Neural Networks》和《软件学报》编委,《Machine Learning》等期刊客座编辑,以及ACML 17、PRICAI 16、CCFAI 17程序主席,AAAI 17、IJCAI 17、ICDM 17等国际会议领域主席或高级程序委员。获NSFC优秀青年科学基金(2012年度)、入选教育部“长江学者奖励计划”青年学者(2015年度)等。

文字内容
1. CCAI人工智能青年论坛 Learning with Weak Supervision (弱监督机器学习范式) Min-Ling Zhang (张敏灵) PALM Group, School of Computer Science and Engineering, MOE Key Laboratory of Computer Network & Information Integration, Southeast University, China July 23, Hangzhou
2. Big Data Essential Goal Turn data into information and knowledge, so as to support sound decision making Key Techniques Cloud Computing Crowdsourcing Machine Learning Managing Data Collecting Data Analyzing Data Min-Ling Zhang Learning with Weak Supervision
3. Traditional Supervised Learning object instance label instance label Input Space represented by a single instance (feature vector) characterizing its properties Output Space associated with a single label characterizing its semantics …… …… …… …… Predictive model Min-Ling Zhang Supervised Learning Algorithm Learning with Weak Supervision
4. Basic Assumption: Strong Supervision label supervision information Key factor for successful learning (encoding semantics and regularities for the learning problem) Strong supervision assumption  Sufficient labeling abundant labeled training data are available  Explicit labeling object labeling is unique and unambiguous Min-Ling Zhang Learning with Weak Supervision
5. But, Supervision Is Usually Weak Difficult to have! Strong supervision (sufficient & explicit) Constrained by:  Limited resources  Physical environment  Problem properties  …… Strong generalization ability In practice, we usually have to learn with weak supervision Min-Ling Zhang Learning with Weak Supervision
6. Learning with Weak Supervision  Insufficient labeling Labeled Data + Unlabeled Data  Non-Unique labeling Multi-Label Data (labeling with multiple valid labels)  Ambiguous labeling Partial-Label Data (labeling with multiple candidate labels) Min-Ling Zhang Learning with Weak Supervision
7. Semi-Supervised Learning (SSL) SSL Learning System Predictive model Major paradigm in exploiting unlabeled data to improve generalization performance, without human interventions  Generative methods [Miller & Uyar, NIPS’97] [Nigam et al., MLJ00]  S3VMs [Joachims, ICML’99] [Chapelle & Zien, AIStats’05] [Grandvalet & Bengio, NIPS’05]  Graph-based methods [Zhu et al., ICML’03] [Zhou et al, NIPS’04] [Belkin et al., JMLR06]  Disagreement-based methods [Blum & Mitchell, COLT’98] [Zhou & Li, KAIS10] Min-Ling Zhang Learning with Weak Supervision
8. Multi-Label Objects Sports Europe Economics Travel Government …… Multiple labels Min-Ling Zhang Learning with Weak Supervision
9. Multi-Label Learning (MLL) object instance label label label …… …… …… Min-Ling Zhang Multi-Label Learning (MLL) Learning with Weak Supervision
10. Major Challenge of MLL input space features output space label sets The MLL Mapping Exponential number of possible label sets ! q=5  32 label sets q=10  ~1k label sets q=20  ~1M label sets …… Supervision Info.  Individually strong  But, globally weak ! Min-Ling Zhang Learning with Weak Supervision
11. Partial Label Appreciator A Picasso style × Appreciator B Monet style × Appreciator C van Gogh style √ Widely exist in real-world applications  Computer vision [Cour et al., JMLR11] [Tang & Zhang, AAAI’17]  Image classification [Zeng et al., CVPR’13] [Chen et al., CVPR’13]  Learning from crowds [Raykar et al., JMLR10] [Yu & Zhang, MLJ17]  Ecoinformatics [Liu & Dietterich, NIPS’12] [Zhang & Yu, IJCAI’15]  …… Min-Ling Zhang Learning with Weak Supervision
12. Partial-Label Learning (PLL) object instance …… …… …… label label label  Each object is associated with multiple candidate labels  Only one of the candidate label is the unknown ground-truth label Partial-Label Learning (PLL) Min-Ling Zhang Learning with Weak Supervision
13. Other Scenarios Widely Exist multi-instance learning [Dietterich et al., AIJ97] [Foulds & Frank, KER10] [Amores, AIJ13] PU learning [Liu et al., ICML’02] [Liu et al., ICDM’03] [Li et al., ACL’10] learning with constraints [Wagstaff et al., ICML’01] [Basu et al., CRCBook08] ………… instance instance must-link can’t-link instance Min-Ling Zhang Learning with Weak Supervision ambiguous labeling insufficient labeling non-unique labeling
14. Other Scenarios Widely Exist multi-instance learning [Dietterich et al., AIJ97] [Foulds & Frank, KER10] [Amores, AIJ13] ambiguous labeling LPUealearrnninigng with Weak Supinseufrfivcieinstilaobneling [Liu et al., ICML’02] [Liu et al., Framework ICDM’03] [Li et al., ACL’10] + Model + Utilization learning with constraints instance instance non-unique labeling [Wagstaff et al., ICML’01] [Basu et al., CRCBook08] must-link can’t-link instance ………… Min-Ling Zhang Learning with Weak Supervision
15. Thanks! Min-Ling Zhang Learning with Weak Supervision