AI - 它可以（和不能）在研發中給你帶來競爭優勢的地方.pdf

《AI - 它可以（和不能）在研發中給你帶來競爭優勢的地方.pdf》由會員分享，可在線閱讀，更多相關《AI - 它可以（和不能）在研發中給你帶來競爭優勢的地方.pdf（37頁珍藏版）》請在三個皮匠報告上搜索。

1、AI-Where it Can(and Cant)Give You a Competitive Edge in R&DThursday,September 05,2024Presented byNoel Hollingsworth|CEO&Co-Founder|Uncountable01|Introduction02|The Current Landscape:AI in R&D03|Considerations Before Implementing AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastruct

2、ure06|The Benefits of an All-in-One Data Infrastructure07|Q&AAgendaFeatured Speaker:Noel HollingsworthCEO&Co-FounderNoel HollingsworthNoel Hollingsworth is Co-Founder and CEO at Uncountable.In his role,he works closely with Uncountables customers to implement next-generation data management systems.

3、Prior to his work at Uncountable,Noel led data teams at startups and was awarded Forbes 30 under 30 for his work with machine learning and artificial intelligence.01|Introduction|Featured SpeakerFounded in 2016 with offices in San Francisco,New York City,and Munich 100+Customers Across Industries R&

4、D organizations including:paints&coatings,cosmetics&personal care,advanced materials,food&beverage,biotechnology&life sciencesOne-of-a-Kind Platform that centralizes R&D data and helps reduce new product development timelinesProven Domain Expertise Began as a data science company helping Fortune 500

5、 materials companies accelerate development of new projects.About Uncountable01|Introduction|About UncountableScan QR Code To Read Customer Case StudiesUncountable Proudly Supports Clients That Span Across a Variety of Industries01|Introduction02|Overview:AI in R&D03|Considerations Before Implementi

6、ng AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastructure06|The Benefits of an All-in-One Data Infrastructure07|Q&ADeep LearningLarge neural networks,taking advantage of large datasets to make predictionsDemonstrated revolutionary success in areas like image recognition and fraud

7、 detection in 2010sGenerative AIPopularized by ChatGPT in 2022Capable of responding to text queries,summarizing data,creating images in“intelligent”mannerTech Companies Rushing to Take Advantage of This1 Trillion in CapEx over coming years02|Overview;AI in R&DAI Today:What Is Causing All The Hype?02

8、|Overview;AI in R&DBut Also Doubts?02|Overview;AI in R&DWhat Characterizes These New Models?CharacterizationFit for R&D?Huge Data Sources NeededIf we can simulate experimental process-yes,otherwise noLarge Amount of Compute Required to TrainYesBlack Box ExplanationsIt dependsResults are not“Perfect”

9、YesResults Reflect DatasetYes,but its a limitationThere will be places where these methods are revolutionary!Large amounts of pre-existing dataPlaces where we can run simulations or do extremely high throughput testingAdjacent areas to experimentation,where we can collect data and learn insightsThey

10、 are not a fit today for many scientific workflowsSmall data availableSimulation not possibleWant reasonable explanations for model behavior02|Overview;AI in R&DWhat Does This Imply?02|Overview;AI in R&DWhere Do We Go From Here?“We dont have better algorithms.We have more data.”“More data beats clev

11、er algorithms,but better data beats more data.”“Simple models and a lot of data trump more elaborate models based on fewer data.”3 Quotes from Peter Norvig(Google Research Director)02|Overview;AI in R&DWhere do we go from here?Get more dataMake sure data is“better”Apply the right models to that data

12、Understand that coatings R&D has inherent challenges:Could prevent an“AI revolution”taking place in an area with more data(e.g.,drug discovery)01|Introduction02|Overview:AI in R&D03|Considerations Before Implementing AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastructure06|The Be

13、nefits of an All-in-One Data Infrastructure07|Q&AThe Core Issue in R&D Orgs:Decentralized,Unstructured,&Fragmented DataData is collected&sits independently across the different teams and systems used throughout the entire R&D value chain Raw MaterialsPhysical AttributesChemical AttributesBatch Histo

14、ryPricesRegulatoryCompliance Compositions Calculations Multi-part systems Order of Addition Dispersions&IntermediatesFormulations Mixing Conditions Curing Parameters MetadataProcessing Physical Properties Analytical Measurements Observations Images Curves&ReportsFormula Properties Application Testin

15、g Customer Feedback Panel TestingProduct Properties03|Considerations Before Implementing AITypes of Data Systems:R&D OrganizationsSpreadsheetsELN/Lab JournalsPredictive toolsVisualizations/AnalysisLIMSOther Internal DatabasesInventoryStatistical Tools03|Considerations Before Implementing AIUnstructu

16、redStructured Spreadsheets Word Documents PDFs Lab Journals/ELNs SharePoint/Shared DriveExamplesData Systems:Structured vs.Unstructured Free Unrestricted entry of information Known/second nature“habitual”Advantages Limited scope&scalability for application of info Ctrl+F keyword searching Limited co

17、llaboration Inability to innovate efficiently&at market-rateDisadvantages Databases LIMS Inventory Systems UncountableExamples Instant access to specific information/data Shareable&scalable information Intelligent insights&reportingAdvantages Requires intentional/deliberate entry of information Chan

18、ge management Migration of historical data into new system Disciplined use Disadvantages03|Considerations Before Implementing AIWhy Excel&Unstructured Data System Are Insufficient1.Volume of DataA small data set with the best AI model in the world is worse than both expert scientists and simpler AI

19、models applied to“big data”The most important aspect of any AI model is its underlying data-both size and cleanliness2.Relevancy to ProblemsWill create desire to squeeze square peg in round hole-When we do have some data,we must apply AI,even if its not a fitAI is not a fit for all use cases!3.Scien

20、tist TrustDesire to be AI-first company without gathering appropriate data=scientist trust being lostAI ends up being applied to projects that arent good fits,or only to high priority projects that carry substantial failure risk-when there are issues,team loses faith in the processSufficient Data is

21、 important,but not the only prerequisiteTop 3 Problems:Deploying AI Without Structured Data03|Considerations Before Implementing AIStandard Way Data Gets Recorded In Spreadsheets&Notebooks:Viscosity,7D=3000Brookfield Visc.Sp#4=5500BV,ON=1800Best Practices for Structuring Lab Data for AI:Brookfield V

22、iscosity=5000Liquid Aging Time+Temperature:7D at 23CSpindle#4RPM:150Test Temperature:23-Exact temperature and time-Machine SN,OperatorImportance of Structuring Lab Data for AI:Example of Brookfield Viscosity03|Considerations Before Implementing AI01|Introduction02|Overview:AI in R&D03|Considerations

23、 Before Implementing AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastructure06|The Benefits of an All-in-One Data Infrastructure07|Q&ACan we just pick the one with the best performance on our test set?Its not that easy!04|Understanding The Resources&Constraints for AI How Do We Se

24、lect The Right Models To Use?04|Understanding The Resources&Constraints for AI Example:Gradient Boosted TreesOften has best performance predicting resultsDecisions made by answering:“Yes”“No”Questions&Summing ResultsImagine this applied to a viscosity predictionIf water water 50%Viscosity is 110 cps

25、Even if results are goodDoes not reflect realityProduces downstream issues04|Understanding The Resources&Constraints for AI Example:Exploitation vs.ExplorationOnce you have a model,you have to decide which points to testWe could test the point our model thinks will perform bestWhat if this is just a

26、 copy of our previous best resultIncidentally,what does“perform best”mean?We could test the point that gives us the most new informationWhat if this is something completely impractical to formulate?Balancing between the two is a key goal in the field of Bayesian OptimizationIn our case,often done wi

27、th many different physical,cost and regulatory constraintsSelecting a data point is significantly more difficult than predicting its performance04|Understanding The Resources&Constraints for AI So,What Do We Do?Dont just assume that good model performance=Good suggestions of experiments to runPerfor

28、mance on a toy example does not reflect performance on your dataIf you implement AI with a partner,emphasize:Is AI something they deeply understand or a trend they are chasing?Anyone can hire someone to be their“AI expert”-Was the company built with AI in mind,or was it added because of where we are

29、 in the hype cycle?Are they trying to sell you on AI for all use cases,or just where its a fit?Are they demanding you spend large amounts of money on nebulous AI projects before seeing results,or do they work with you to take it step by step?01|Introduction02|Overview:AI in R&D03|Considerations Befo

30、re Implementing AI04|Understanding The Resources&Constraints for AI05|Best Practices:Start Creating The Proper Infrastructure06|The Benefits of an All-in-One Data Infrastructure07|Q&AToo big of a search space100s of ingredients,but limited data pointsEither from collection,cleanliness,or standardiza

31、tionMoonshot objectivesWhat are you trying to achieve in this project vs long term goalsWhat are more reasonable targets that would allow you to claim“progress”Perception of perfectionWhy would model suggest such a thing?Why isnt model more accurate?Can it model pictures of exposure ratings?Consider

32、ations:Setting The Right Expectations05|Best Practices:Creating The Proper Infrastructure for AI1.Before(Preparation)Ensure structured data system in placeVerify all scientist work is being captured in a way fit for AIAll data points and all aspects of dataExample:Viscosity centipoise,temperature,sp

33、indle,rpmUtilize in-house expertise to understand/validate vendor and partner“claims”2.During(Deployment)Identify appropriate targets for AI-Example Criteria:Large Amounts of DataKnown Success CriteriaConsistent Output ResultsEnsure AI is embedded into daily workflowsNot judged off success in a proj

34、ect where majority of results are out of scientists control3.After(Maintenance)Identify areas where data capture is insufficientDeploy systems and/or recurring procedures to collect dataHow to Define&Create an AI Roadmap05|Best Practices:Creating The Proper Infrastructure for AI01|Introduction02|Ove

35、rview:AI in R&D03|Considerations Before Implementing AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastructure06|The Benefits of an All-in-One Data Infrastructure07|Q&AUncountables All-In-One Platform?A Platform to Centralize,Connect,And Structure All Types Of R&D Data.06|The Benefi

36、ts of an All-in-One Structured Data System New Modern Digitalization Tools:R&D Labs in Paints&CoatingsDrive learnings from the raw material level to improve performanceBetter understanding of the correlation between the data specific to raw materials and the desired results for formulas Integration

37、with raw materials/costs systemsEnable world where scientists dont have to go to multiple systems to formulate with context Unified Laboratory Informatics PlatformsEnd-to-end web-based systems connecting recipes to resultsData is collected in standardized,streamlined,consistent wayAI/ML-powered DOE

38、approach to efficiently explore the defined formulation spaceEnabling acceleration towards sustainability goals06|The Benefits of an All-in-One Structured Data System Example 1Platform-Wide Data06|The Benefits of an All-in-One Structured Data System Example 1:Output Fits06|The Benefits of an All-in-

39、One Structured Data System Example 1:Linear Coefficients06|The Benefits of an All-in-One Structured Data System Example 2Targeted Experiments06|The Benefits of an All-in-One Structured Data System Example 2:Suggested Experiments06|The Benefits of an All-in-One Structured Data System 01|Introduction0

40、2|Overview:AI in R&D03|Considerations Before Implementing AI04|AI Modeling05|Best Practices:Start Creating The Proper Infrastructure06|The Benefits of an All-in-One Data Infrastructure07|Q&AJust a few of manyUncountable Long-Term CustomersQ&AThank You!Questions?Email: Inquiries: a Demo?Scan The QR Code