LytleLecture2023-2024Objective-DrivenAITowardsAIsystemsthatcanlearn,remember,reason,plan,havecommonsense,yetaresteerableandsafeYannLeCunNewYorkUniversityMeta–FundamentalAIResearchUniversityofWashingtonLytleLecture2024-01-24M-51,HSOMachineLearningsucks!(comparedtohumansandanimals)Y.LeCunSupervisedlearning(SL)requireslargenumbersoflabeledsamples.Reinforcementlearning(RL)requiresinsaneamountsoftrials.Self-SupervisedLearning(SSL)worksgreatbut...GenerativepredictiononlyworksfortextandotherdiscretemodalitiesAnimalsandhumans:Canlearnnewtasksveryquickly.UnderstandhowtheworldworksCanreasonanplanHumansandanimalshavecommonsenseTherebehaviorisdrivenbyobjectives(driv...