AI Full.pptx [Read-Only] - DARPA

1 Approved for Public Release, Distribution , complex and subtle informationlearnwithin an environmentabstractto create new meaningsreasonto plan and to decideArtificial intelligence is a programmedability to process informationperceivingabstractingreasonin glearningNotional intelligence scaleApproved for Public Release, Distribution KnowledgeStatistical LearningContextual AdaptationApproved for Public Release, Distribution KnowledgeApproved for Public Release, Distribution reasoning over narrowly defined problemsNo learning capabilityand poor handling of uncertaintyPerceivingAbstractingReasonin gLearningApproved for Public Release, Distribution create sets of rules to represent knowledge in well defined domainsThe structureof the knowledge is defined by humansThe specificsare explored by the machineApproved for Public Release, Distribution Autonomous Vehicle Grand Challenge140 miles of dirt tracks in California and Nevada 2004# completed: 02005# completed: 5 Source: DARPAA pproved for Public Release, Distribution LearningSource: for Public Release, Distribution create statistical models for specific problem domains and train them on big dataSource: for Public Release, Distribution classification and prediction capabilitiesNo contextual capability and minimal reasoning abilityPerceivingAbstractingReasoningLea rningApproved for Public Release, Distribution hypothesisNatural data forms lower dimensional structures (manifolds) in the embedding spaceApproved for Public Release, Distribution data comes by separating the manifoldsEach manifold represents a different entityApproved for Public Release, Distribution ,253,247,228,118,41,38,34,39,147,198.

2 Approved for Public Release, Distribution in handwritten digits form 10 distinct manifolds within the28x28 dimensional space of pixel valuesApproved for Public Release, Distribution the spiral arms are each clusters of dataStretching and squashing the data space separates them cleanlyApproved for Public Release, Distribution in a new dimension enables enclosed manifoldsto be isolatedApproved for Public Release, Distribution layer stretches and squashes the data space until the data manifolds are cleanly separatedApproved for Public Release, Distribution POS (SUMPRODUCT( W1:W16, V1:V16))cell weights (learned)cell inputs(from prev layer)non linear functionApproved for Public Release, Distribution feature map performs a local analysis over the whole input spaceFully connected layers perform global analysis1000fullyconnected20 feature mapseach 24x24convolutionssubsamplingconvolutions subsamplinginput28x28018912x12feature mapsapprox.

3 30,000cells in total for> accuracyMachine learning programmers design the network structure with experience and by trial and errorApproved for Public Release, Distribution for Public Release, Distribution group of people shopping at an outdoor marketThere are many vegetables at the fruit standYann LeCun, Yoshua Bengio, & Geoffrey Hinton (2015). Deep Learning, Nature, Vol. 521, (pp. 436 444)A deep convolution neural net (CNN) produces a set of outputs (abstract words )A language generating recurrent neural net (RNN) translates the abstract words into captionsApproved for Public Release, Distribution and network flowsObserve real time cyber attacks at scaleElectromagnetic spectrumOvercome spectrum scarcity to meet wireless data demandAutonomous platformsReshape defense missionsApproved for Public Release, Distribution young boy is holding a baseball bat Statistically impressive, but individually unreliableApproved for Public Release, Distribution Panda Gibbon ( confidence)

4 +=< 1% targeted distortionInherent flaws can be exploitedApproved for Public Release, Distribution training data creates maladaptationInternet trolls cause the AI bot, Tay, to act offensivelyApproved for Public Release, Distribution adaptationSystems construct contextual explanatory modelsfor classes of real world phenomenaApproved for Public Release, Distribution DataExplainable ModelExplanationInterfaceThis is a cat: It has fur, whiskers, and claws. It has this feature: mistakeI understand whyI understand why notI know when you ll succeedI know when you ll failI know when to trust youI know why you made that mistakeSource: SPIN South WestApproved for Public Release, Distribution dataGenerative modelGenerates explanations of how a test character might have been createdProbable number of strokes: 1 4 Each stroke: probable trajectoryEach trajectory: probable shift inshape and locationSeed modelApproved for Public Release, Distribution modelperceiveApproved for Public Release, Distribution KnowledgeStatistical LearningContextual Adaptatio