WIXLIB

Artificial Intelligence: LearningThrough Interactions and Big DataAn IBM Redbooks Point-of-View publication by IBMAcademy of TechnologyAuthored by Koichi Takeda, IBM ResearchTokyo, and Tamiya Onodera, IBM ResearchTokyoHighlightsAI approaches combined with analyticstechniques can provide practicalapproaches to the challenges of creatingsmarter solutions. The current mainstream businessanalytics research and solutiondevelopment along with theconvergence of AI with machinelearning techniques will continue tounderpin high-value, human decisionsupport solutions. Cognitive systems can combine naturallanguage processing, hypothesisgeneration and evaluation, and dynamiclearning for a powerful, fast, andintelligent problem solving. To effectively apply intelligent problemsolving using AI solutions, you canextend the capabilities of IBM Watsonand its DeepQA technology architectureto other domains.Artificial intelligence (AI) has been formed as a field of computerscience for understanding and building intelligent entities sincethe 1950s. As with most technologies, the concept hasundergone many transformations over time. In particular, thefield of AI has experienced several cycles of expectations,commitments, and then disappointments, sometimes referred toas AI winters, where interest in the technology has waned onlyto then peak again at a later time. Some might argue that thesecycles of conflict have caused inconsistencies that have madeforward progress difficult in the AI field. However, recentpractices of data-centric approaches (for example, machinelearning and probabilistic reasoning) are beginning todemonstrate the true capability of AI to solve practicaldecision-making problems.The concept of AI is often thought of as having two dimensions: One dimension emulates the human thought process orsimulates intelligent behavior (such as cognitive computingand brain-inspired systems). The other dimension achieves human or ideal performance(such as question-answering and natural languageunderstanding capability as demonstrated by IBM Watsonand its Deep Question-Answering technology).This dual nature of AI challenges organizations with how to bestapproach AI modeling solutions. Conventional AI approachesoften tackle knowledge acquisition and reasoning in terms of aset of ad hoc rules and manually crafted weights for computingconfidence. As a result, acquiring and maintaining consistencyof rules and improving the accuracy of the reasoning can be achallenge when the size of the knowledge base grows to solvereal-world problems. This knowledge bottleneck for expertsystems is often perceived as an inhibitor for building scalableindustrial solutions.However, AI approaches remain central to intelligent problemsolving. Unlike precisely defined mathematical problems, wherealgorithms for computing complete or approximate answers canbe applied, many real-world problems are often underspecified,ambiguous, or incompletely defined and understood. AItechniques turn out to be effective for modeling the problem andbuilding practical solutions.One important aspect of modern AI is that it now embracesmachine learning techniques for building statistical models andmaking statistical reasoning and prediction. Recent advances instatistical approaches along with the availability of a massiveamount of data (big data) have been remarkable. Theseapproaches are driving enhancements for high-value businessanalytics.Redbooks Copyright IBM Corp. 2013.1

According to Bernard Meyerson, Chief InnovationOfficer, IBM, the DeepQA technology of IBM Watson “isjust a first step into a new era of computing that’s goingto produce machines that are as distinct from today’scomputers as those computers are from themechanical tabulating devices that preceded them.”1As organizations embrace AI, understanding themethods to approaching intelligent problem solving iscentral to gaining an advantage with AI solutions.AI and business analytics drivesmarter solutionsMachine learning and statistical or predictive analysis,in many cases, boils down to business analytics, whichis a core component of contemporary AI. Thus,business analytics are central to the AI-relatedactivities that build a smarter solution. When youapproach methods to use AI technologies, consider thefollowing AI-related activities:doctor can better determine possibilities for treatment.In turn, better data might help reduce unnecessarytreatments or tests, thus improving patient outcomesand reducing overall costs, for the patient, the practice,and ultimately, the healthcare industry, while alsomaking processes more efficient.In another example, a cognitive computing assistant,such as IBM Watson Engagement Advisor, can help abusiness to better serve its consumers. Thistechnology allows brands to crunch big data andtransform the way that they engage clients in keyfunctions, such as customer service, marketing, andsales. IBM Watson Engagement Advisor can learn,adapt, and understand a company’s data, enablingcustomer-facing personnel to assist consumers withdeeper insights more quickly than previously possible.Other examples of adapting AI technologies to otherindustries might be the deployment of speechrecognition and voice command interpreters tosmartphones, consumer electronics, and automotivesand the application of computer vision to robots. Adapting AI technologies to other industries Applying cognitive computing to mind-likecomputational models Converging AI and business analytics Incorporating operational intelligence into expertintegrated systemsAdapting AI technologies to otherindustriesAn AI program can use natural language capabilities,hypothesis generation, and evidence-based learning toproduce a better outcome. Researchers are nowlooking at advances in technologies to leverage AI inother industries. Facts strongly suggest that thequestion and answer architecture behind the AItechnology that made IBM Watson possible can beapplied to other domains. If quality information sourcesare available and a reasonable set of sample questionsand answers for the domain exists and if the deeplinguistic analysis for the source language is available,this type of technology can adapt well to otherindustries.For example, the core technology behind AI can helptransform the way a doctor interacts with a patient. Byfeeding data to an AI program, such as IBM Watson, ange-the-world-within-five-years.htmlA system that generates hypotheses canrecognize different probabilities of variousoutcomes. It can “learn” from successes andfailures and can improve future responses.Applying cognitive computing tomind-like computational modelsCognitive computers move beyond simple processingof data and begin to learn through experiences. Byapplying cognitive computing, models can findcorrelations, can create hypotheses, and then learnfrom the outcomes. This process is fundamentalresearch for building the coherent and unified model ofmind-like computation. Potentially, the resulting modelcan be applied to solve massive sensory informationprocessing and cognitive problems.Researchers at IBM have been working on a cognitivecomputing project called Systems of NeuromorphicAdaptive Plastic Scalable Electronics (SyNAPSE). Byreproducing the structure and architecture of thebrain—the way its elements receive sensory input,connect to each other, adapt these connections, andtransmit motor output—the SyNAPSE project modelscomputing systems that emulate the brain’s computingefficiency, size, and power usage without beingprogrammed.Artificial Intelligence: Learning Through Interactions and Big Data

One of the earlier achievements in the SyNAPSEproject is the (cat-scale) cortical simulations (with 109neurons and 1,013 synapses) that were run by the BlueGene/P supercomputer with 147,456 CPUs and144 TB of total main memory at the LawrenceLivermore National Laboratory.Cognitive systems can transform howorganizations think, act, and operate.Learning through interactions, they deliverevidence-based responses that can drive betteroutcomes.Converging AI and business analyticsRecent empirical AI approaches deploy analyticalmethods—called machine learning—by assuming astatistical model for capturing uncertainty in the givenproblem. Machine learning works by clearly definingbusiness goals, exploring data, and validating anddeploying models. It relies on domain and dataexpertise, rather than analytical skills. This methodestimates a set of parameters of a model (theprobability) from an ample amount of observed datathat approximates the problem space.Machine learning techniques are often theoreticallysound, and various mathematical properties are knownand used to formulate the given problem. Machinelearning approaches are inspired by the advances andsuccess of Business Analytics and MathematicalScience (BAMS) research. BAMS conducts activeresearch in several areas of analytical andmathematical sciences such as optimization,algorithms, stochastic analysis, statistics, data mining,predictive modeling, and simulation.BAMS current projects include a diverse range ofindustry domains, such as retail and banking servicelines, including Supply Chain Management (SCM) andHuman Capital Management (HCM) and solutionareas, such as Smarter Cities and Smarter Energy.Examples of projects that BAMS has executed includecapacity utilization, hiring and training, strategicplanning and budgeting, workforce scheduling,contingency planning, dynamic supply chainoptimization, customer targeting, smarter water supplymanagement for growth markets, propensity modelingfor cross-sell and up-sell, and so forth.For more information about BAMS, vergence of AI and business analytics approachescan further extend the scope of analytics intostructured and unstructured information, humanbehavioral patterns, and interactive learning ofknowledge. The convergence of AI and analyticspromises to model intelligent decision processes.Incorporating operational intelligenceinto expert integrated systemsToday’s applications are complex. They involve adiverse set of hardware and software components thatinteract with each other horizontally and vertically.Optimizing, deploying, configuring, and managingthese components require an enormous amount ofmanual effort by human experts. This tuning process isextremely laborious and error-prone. Automating thisprocess in a cost-effective manner belongs in the realmof applications for artificial intelligence.In the mid 2000s, IBM researchers successfully appliedcontrol theory to software systems, resulting in two keyfeatures in a commercial product: Utility Throttling of DB2 v8.2 Adaptive Self-Tuning Memory of DB2 v9These features made it easier to configure andmaintain databases for all market segments.In mid 2012, IBM delivered PureSystems, a breed ofexpert integrated systems that capture and automatewhat experts do in order to make every part of the ITlifecycle easier. PureSystems uses a semantically richlanguage, developed by IBM Research, to define itsbuilding blocks. The knowledge that formerly sprawledacross hundreds of pages of manuals is now distilledinto a formal representation.As the technology integrates further, expert integratedsystems will evolve by incorporating more usage datafor improving performance and consumability, whichwill eventually be known as operational intelligence.A grocer who is stocking store shelves can usean instrumented glove designed for a cognitivecomputing system that monitors sights, smells,texture, and temperature to flag bad orcontaminated produce.3