釋放AI伺服器與機器學習潛力李永輝IBM大中華區硬體系統部首席技術官暨傑出工程師勢不可擋的AI市場“By2020,80%ofBigDataandAnalyticsdeploymentswillneeddistributedmicroanalyticsand40%ofallbusinessanalyticssoftwarewillincorporateprescriptiveanalyticsbuiltoncognitivecomputingfunctionality.BothofthesetrendsrequireadramaticincreaseinprocessingpowerthatcouldbeenabledbyGPUs.”—IDC3DataComputeAlgorithms組成AI的三個元件各行各業的數據變換INCREASINGDATAVARIETYSearchMarketingBehavioralTargetingDynamicFunnelsUserGeneratedContentMobileWebSMS/MMSSentimentHDVideoSpeechToTextProduct/ServiceLogsSocialNetworkBusinessDataFeedsUserClickStreamSensorsInfotainmentSystemsWearableDevicesCyberSecurityLogsConnectedVehiclesMachineDataIoTDataDynamicPricingPaymentRecordPurchaseDetailPurchaseRecordSupportContactsSegmentationOfferDetailsWebLogsOfferHistoryA/BTestingBUSINESSPROCESSPETABYTESTERABYTESGIGABYTESEXABYTESZETTABYTESStreamingVideoNaturalLanguageProcessingWEBDIGITALAI釋放Data的價值LanguageUnderstandingLanguageTranslationSpeechTranscriptionFaceRecognitionMachineReasoningObjectDetectionknowledge科技案例進展AI訓練的數據處理DefinetrainingtaskPreparetrainingDataDataPre-processingDNNModelselectionConfigurethetraininghyper-parameterDNNModelTrainingStartPackagethenewDNNmodeltogetherwithpreprocessingintoinferenceproc.DLtrainingframeworkpreparationApplicationdevelopmentwithinferenceAPITobuildateamwithdeeplearningexpertise:2months~1yearTopreparemassivetrainingdata:~10manmonth(s)Totrainanewmodel:1hour~weekTogiveanewinferenceresult:<1secondCognitiveSystem:ProvideoptimizedSW+HWdesign,andtoolchainstosignificantlyenhance•Productivity•Performance•Timetomarket具備URLI的智能系統LearningDeepLearning/MachineLearningImage/Video/VoiceRecognitionReasoningHighPerformanceDataAnalyticsKnowledgeRepresentationandReasoningUnderstandNaturalLanguageProcessing(NLP)UnstructuredInformationManagementInteractiveTextToSpeech(TTS)/SpeechtoText(STT)QuestionAnsweringTechnologyDataComputeAlgorithms組成AI的三個元件邁向AI之路Z14伺服器L...
