AI: Expectations and RealitiesHard problems in AI -- Machine learning: fragility, opacity, and dynamism... • How do we engineer systems to safely deliver AI to the mission? • How do we harmonize domain models with AI techniques? • How can humans best partner with AI-enhanced systems? • What are successful models for continuous delivery, continuous integration and continuous verification for AI? • What is next generation AI beyond symbolic and machine learning (waves 1 and 2 have shown their limits)?Information Innovation OfficeI2O_64fb56b0-7235-11ea-870b-c6940183ea00* Advantage in cyber operations * Artificial intelligence to the mission * Resilient, adaptable, and secure systems * Confidence in the information domainDARPADARPA Achievements: * MATERIAL SCIENCE *GPS * NIGHT VISION * PREDATOR AUTONOMOUS VEHICLE * NEUROPROSTHETIC LIMBS * STEALTH FIGHTER * ADVANCED SEMICONDUCTORS * SATURN V * AUTONOMOUS VEHICLES * PERSONALIZED ASSISTANT THAT LEARNSDr. William ScherlisDirector, Information Innovation OfficeI2O ProgramsCommunicating with Computers (CwC) ProgramComputers and Humans Exploring Software Security (CHESS) ProgramExplainable AI (XAI) ProgramLearning with Less Labeling (LwLL) ProgramMachine Common Sense (MCS) ProgramSynergistic Discovery and Design (SD2) Program3G AI_64fb58cc-7235-11ea-870b-c6940183ea00To solve hard problems in AI_64fb5976-7235-11ea-870b-c6940183ea00ResilienceResilient, adaptable, and secure systemsAdaptabilityOpennessOpen – Hardware/software decouplingProgrammabilityProgrammable – Configure to the missionSecuritySecure – Trust and securityAI CoreDevelop core AI_64fb5a48-7235-11ea-870b-c6940183ea001Machine Common Sense (MCS) ProgramContext: Specialized cognitive building blocks: perception, reasoning, action. Approach: • Hybrids methods • Machine learning + game theory + optimization • Machine learning + explicit reasoning • Infrastructure: Computing and data handling • Looking ahead: Self adaptation – learning to learn ^ Frame specialized AI using common sense reasoning.Situational UnderstandingEnable AI applications to understand new situations_64fb5ad4-7235-11ea-870b-c6940183ea001.1ReasonablenessEnable AI applications to monitor the reasonableness of their actions_64fb5b60-7235-11ea-870b-c6940183ea001.2Learning TransferEnable AI applications to transfer learning to new domains_64fb5bec-7235-11ea-870b-c6940183ea001.3CommunicationEnable AI applications to communicate more effectively with people_64fb5c78-7235-11ea-870b-c6940183ea001.4PeopleApplicationsAdvance mission applications of AI_64fb5cfa-7235-11ea-870b-c6940183ea002Knowledge-directed AI Reasoning Over Schemas (KAIROS) ProgramContext: • Emerging AI-enabled mission concepts • Adversaries are nimble and capable • Human-AI partnering remains difficult • Talent pool is a challenge ^ Approach: • Close partnering of operators and engineers • Start with advisory AI ReasoningCreate schema-based artificial intelligence capability to enable contextual and temporal reasoning about complex real-world events_64fb5d90-7235-11ea-870b-c6940183ea002.1Text | Speech | Images | Video > Media Analysis with Temporal Annotation > Schema Application & Temporal Reasoning > Temporal Knowledge Base > Predictive Analysis > User InteractionSystems EngineeringEngineer systems with embedded AI_64fb5e1c-7235-11ea-870b-c6940183ea003Explainable AI (XAI) ProgramContext: Software and systems engineering are made more challenging with AI. Approach: • Adapt key aspects of the engineering process • Integration frameworks, planning, and design • Process, tooling, and measurement • Assurance and evidence • Data, systems infrastructure, and configurations Understanding, Trust & ManagementEnable human users to understand, trust, and effectively manage the emerging generation of AI partners_64fb5eb2-7235-11ea-870b-c6940183ea003.1HumansAI PartnersExplain second wave AICausal ModelsLearn more structured, interpretable, causal models_64fb5f52-7235-11ea-870b-c6940183ea003.1.1FeaturesLearn more explainable features_64fb61f0-7235-11ea-870b-c6940183ea003.1.2Black-Box ModelsInfer an explainable model from any model as a black-box_64fb629a-7235-11ea-870b-c6940183ea003.1.3Evaluation & AcceptanceDevelop continuous evaluation and acceptance_64fb633a-7235-11ea-870b-c6940183ea004Guaranteeing AI Robustness against Deception (GARD) ProgramContext: • Machine learning fragility, opacity, and dynamism • Adversaries empowered in new ways, including attacking conventional systems • Assurance influences all aspects of engineering and design, from the outset ^ Approach: • Integrate assurance planning • Manage evidence to support confident accreditation decisionsDeceptionEnable machine learning systems to be robust against adversary deception_64fb63da-7235-11ea-870b-c6940183ea004.1Design robust and resilient AI models2020-03-29https://download.1105media.com/Custom/Workshops/2020/AI/Bill_Scherlis.pdfOwenAmburOwen.Ambur@verizon.net