人工智能青年论坛：弱监督机器学习范式——张敏灵 东南大学教授（15页）.pdf

1、Learning with Weak Supervision（弱监督机器学习范式）（弱监督机器学习范式）CCAI人工智能青年论坛July 23,HangzhouPALM Group,School of Computer Science and Engineering,MOE Key Laboratory of Computer Network&Information Integration,Southeast University,ChinaMin-Ling Zhang(张敏灵张敏灵)Min-Ling ZhangLearning with Weak SupervisionBig DataEss

2、ential GoalTurn data into information and knowledge,so as to support sound decision makingKey TechniquesCloud ComputingManaging DataCrowdsourcingCollecting DataMachine LearningAnalyzing DataMin-Ling ZhangLearning with Weak SupervisionTraditional Supervised LearningobjectinstancelabelInput Spacerepre

3、sented by a single instance(feature vector)characterizing its propertiesOutput Spaceassociated with a single label characterizing its semanticsSupervised Learning AlgorithmPredictive modelinstancelabelMin-Ling ZhangLearning with Weak SupervisionBasic Assumption:Strong Supervisionlabelsupervision inf

4、ormationKey factor for successful learning(encoding semantics and regularities for the learning problem)Strong supervision assumption Sufficient labelingabundant labeled training data are available Explicit labelingobject labeling is unique and unambiguousMin-Ling ZhangLearning with Weak Supervision

5、But,Supervision Is Usually WeakStrong supervision(sufficient&explicit)Strong generalization abilityDifficult to have!Constrained by:Limited resources Physical environment Problem properties In practice,we usually have to learn with weak supervisionMin-Ling ZhangLearning with Weak SupervisionLearning

6、 with Weak Supervision Insufficient labelingLabeled Data+Unlabeled Data Non-Unique labelingMulti-Label Data(labeling with multiple valid labels)Ambiguous labelingPartial-Label Data(labeling with multiple candidate labels)Min-Ling ZhangLearning with Weak SupervisionSemi-Supervised Learning(SSL)Predic