Upload File

cscibox: an intelligent assistant for dang ice and sediment cores …lizb/cscience/agu2015.pdf ·...

  • Home
  • Documents
  • CSciBox: An intelligent assistant for dang ice and sediment cores …lizb/cscience/agu2015.pdf · 2016-07-09 · AGU2015-CSciBox-Poster.pptx Author: Liz Bradley Created Date: 7/9/2016
1
AI-geoscien+st interac+on • The scien+st is kept “in the loop” at each decision point, and can guide the process as she sees fit. At each step, she can compare the results to her intui+on or to other age models. • Hobbes explores alterna+ves, performs computa+ons, and tracks uncertainty. It explores the space of possible models thoroughly and quickly, and always explains its reasoning to the scien+st, including steps that are conflic+ng, uncertain, or otherwise problema+c. Hobbes’s Knowledge Base Hobbes is aiming to encode the vast knowledge geoscien+sts have built up through years of experience da+ng cores. Hobbes’s collec+on of rules springs out of long, detailed discussions between the AI scien+sts and the geoscien+sts on the team. These conversa+ons focus on 1. The methods of reasoning geoscien+sts use to create & evaluate age models 2. Specific rules of inference geoscien+sts use in par+cular situa+ons while crea+ng an age model Examples of rules obtained from these conversa+ons include: • The more measurements we have of a core, the more certain we’ll be about the resul+ng age model. • Stra+Counter works well for layer coun+ng if neighboring layers have roughly the same thickness. • If a sediment layer contains fossils of terrestrial life, it must have been deposited on land. • If the sedimenta+on rate changes abruptly, the age model is probably wrong. The AI scien+sts then encode each rule of inference into an algorithm Hobbes can understand. Evidence FOR snow melt Temperature frequently above freezing (stronger evidence) Evidence AGAINST snow melt Average temperature below freezing (weaker evidence) Example: Da+ng NEEM-S1 ice core Step 1: Use Stra:Counter [Winstrup 15] for automated layer coun:ng Hobbes: I need a parameter \mu, for the average layer thickness. I guessed \mu= 7 cm, based on a similar core. Do these layer boundaries look correct? Geocien+st: No, there are too many layer boundaries. Use \mu = 13 cm. Hobbes: Okay, using \mu = 13 cm. Do these layer boundaries look correct now? Geocien+st: Yes, con+nue layer coun+ng. Step 2: Align the gas record and the ice record Hobbes: We need a firn densifica+on model to es+mate age difference between gas bubbles and surrounding ice. The standard model is by Herron&Langway, but it assumes there was no snow melt. I will evaluate whether this model is appropriate. Argument FOR snow melt (only moderately certain) Hobbes: Judging by the current temperature values at the loca+on of the NEEM-S1 core (average -3.34 C and standard devia+on 22.97), the Herron&Langway model may not be appropriate for this core. Should I use the <name other model here> instead? Step 3: Stop layer coun:ng aDer annual signals fade Hobbes: We are at depth 3525 m. I will check whether the annual signals have faded to see if I should keep layer coun+ng. Argument: number of peaks per series is normal partly applicable sound Argument: average layer thickness is normal mostly applicable plausible Conclusion: con+nue layer coun+ng mostly applicable sound Hobbes: We can con+nue layer coun+ng ajer depth 3525 m. Final output: Reasoning behind age models: • Performed automated layer coun+ng between depth 50 m and 4020 m <these are made up!> • Used Dansgaard-Johnsen flow model for da+ng depths below 4020 m I don’t see that above. I think you’re missing a 4 th step • Used two different firn densifica+on models: Herron&Langway (orange curve) and ? (blue curve) • What about the gas record alignment? • Etc (more details would be included in the Reproducibility Reproducibility is a key component of the scien+fic method. As computa:on becomes more central to the scien+fic enterprise, it is urgent to address concerns regarding reproducibility of computa:onal results. CSciBox addresses reproducibility concerns by: • Tracks all ac+ons taken with the core, for complete reproducibility • Includes cita+on notes every +me you use another scien+st’s tool • Metadata is always bundled with the core data • We use Linked Paleoclimate Date format • It is easy (and mandatory) in CSciBox to save all the data and metadata that produced an age model (if you want to save the age model). By metadata, we mean specific parameter choices and such- everything that’s required to replicate the computa+on exactly. • It is also easy (but op+onal) to save informa+on about the other models that the scien:st considered but rejected. <Something about why this is important for science too> • Hobbes’s knowledge base encodes the reasoning behind geoscience research. This reasoning may not always be reported in tradi+onal scien+fic journaling (?) The process of “algorithmizing” the rules of inference forces scien+sts to codify and report all of their reasoning. That’s cool! • CSciBox specifically addresses many of the recommenda+ons of the Yale Law School Roundtable on Data and Code Sharing (cite). Argumenta+ve reasoning Hobbes considers arguments for and against each theory, and weighs their strength. • This mimics the way scien:sts think, and the way they communicate about it with others—and thus streamlines communica+on between Hobbes and its users • It also allows the scien+st to consider mul:ple age models at once, and enables Hobbes to handle conflic+ng arguments. • At this point in the process, for instance, the scien+st can choose one or the other of the suggested firn models—or may choose todirect Hobbes to explore both. • This is the place to put that material about scien+fic disagreement and conflic+ng rules Two components of confidence in the reasoning • Confidence in the scien:fic knowledge (validity): Is this rule of inference valid? Validity ranges from a universally accepted rule (accepted), to a gut-level feeling (plausible). Accepted > Sound > Probable > Plausible A validity level is aqached to each rule, and is determined by the expert geoscien:sts on the team. The discrete levels of validity correspond to the weights scien+sts give to their reasoning, and are easier to obtain from experts than con+nuous values would be. • Confidence in the evidence (applicability): How well does our situa+on match the rule of inference, as shown by our dataset? Highly applicable > mostly applicable > partly applicable Hobbes determines the applicability of a piece of evidence, by analyzing the data. • When weighing the evidence for or against a theory, Hobbes is designed to intelligently consider the validity and applicability of each argument.. Hobbes combines the individual arguments into a single conclusion, Hobbes: Okay [saves \mu=13 cm in the metadata and proceeds] PP11A-2203 [email protected] Elizabeth Bradley 1 , Kathleen Finlinson 2 , Tyler Jones 3 , Colin Lindsey 3 , Brett Israelsen 1 Kenneth A. Anderson 1 , Thomas M. Marchitto 3 , Laura Rassbach de Vesine 1 , and James W. C. White 3 , 1 Department of Computer Science, University of Colorado, Boulder, Colorado, USA 2 Department of Applied Mathematics, University of Colorado, Boulder, Colorado, USA 3 Institute for Alpine and Arctic Research (INSTAAR), University of Colorado, Boulder, Colorado, USA CSciBox: An intelligent assistant for da+ng ice and sediment cores The main idea: CSciBox Geoscience Computer science Image courtesy of Dorthe Da Photo: Dorthe Dahl-Jensen Geoscien+st: • Has core • Wants age model • Brings scien+fic exper+se CSciBox Output: • One or more proposed age models for the core • Sophis+cated uncertainty es+mates • Summary of the reasoning behind each model Hobbes CSciBox : • Helps apply scien+fic reasoning and algorithms • Explores alterna+ves automa+cally • Keeps track of all the reasoning involved in the model crea+on • Need to explain the rela+onship between Hobbes and CSciBox This is old; will need a new screenshot This material is based upon work sponsored by the Na+onal Science Founda+on. Any opinions, findings, and conclusions or recommenda+ons expressed in this material are those of the author(s) and do not necessarily reflect the views of the NSF. www.cs.colorado.edu/~lizb/cscience.html • Source code (python) available on github: github.com/ldevesine/cscience • But you don’t have to know python to run it; we also have one-click installers there too • GNU public license; free to modify/extend/use as you see fit We are trying to build— and support — a user community. Please join us! CSciBox is open source In the works: • Say what the future work landscape looks like •Here are some sugges+ons: •Making Stra+Counter & Bacon both work •Helping Hobbes make smart decisions about when to run expensive computa+ons •Adding extensively to Hobbes’ rule base

Upload: others

Post on 15-Jul-2020

0 views

Category:

Documents


0 download

Report

TRANSCRIPT

Page 1: CSciBox: An intelligent assistant for dang ice and sediment cores …lizb/cscience/agu2015.pdf · 2016-07-09 · AGU2015-CSciBox-Poster.pptx Author: Liz Bradley Created Date: 7/9/2016

AI-geoscien+stinterac+on•  Thescien+stiskept“intheloop”ateachdecision

point,andcanguidetheprocessassheseesfit.Ateachstep,shecancomparetheresultstoherintui+onortootheragemodels.

•  Hobbesexploresalterna+ves,performscomputa+ons,andtracksuncertainty.Itexploresthespaceofpossiblemodelsthoroughlyandquickly,andalwaysexplainsitsreasoningtothescien+st,includingstepsthatareconflic+ng,uncertain,orotherwiseproblema+c.

Hobbes’sKnowledgeBaseHobbesisaimingtoencodethevastknowledgegeoscien+stshavebuiltupthroughyearsofexperienceda+ngcores.Hobbes’scollec+onofrulesspringsoutoflong,detaileddiscussionsbetweentheAIscien+stsandthegeoscien+stsontheteam.Theseconversa+onsfocuson1.  Themethodsofreasoninggeoscien+stsusetocreate&

evaluateagemodels2.  Specificrulesofinferencegeoscien+stsuseinpar+cular

situa+onswhilecrea+nganagemodelExamplesofrulesobtainedfromtheseconversa+onsinclude:

•  Themoremeasurementswehaveofacore,themorecertainwe’llbeabouttheresul+ngagemodel.

•  Stra+Counterworkswellforlayercoun+ngifneighboringlayershaveroughlythesamethickness.

•  Ifasedimentlayercontainsfossilsofterrestriallife,itmusthavebeendepositedonland.

•  Ifthesedimenta+onratechangesabruptly,theagemodelisprobablywrong.

TheAIscien+ststhenencodeeachruleofinferenceintoanalgorithmHobbescanunderstand.

EvidenceFORsnowmeltTemperaturefrequentlyabovefreezing(strongerevidence)

EvidenceAGAINSTsnowmeltAveragetemperaturebelowfreezing(weakerevidence)

Example:Da+ngNEEM-S1icecoreStep1:UseStra:Counter[Winstrup15]forautomatedlayercoun:ng

Hobbes:Ineedaparameter\mu,fortheaveragelayerthickness.Iguessed\mu=7cm,basedonasimilarcore.Dotheselayerboundarieslookcorrect?

Geocien+st:No,therearetoomanylayerboundaries.Use\mu=13cm.Hobbes:Okay,using\mu=13cm.Dotheselayerboundarieslookcorrectnow?

Geocien+st:Yes,con+nuelayercoun+ng.

Step2:AlignthegasrecordandtheicerecordHobbes:Weneedafirndensifica+onmodeltoes+mateagedifferencebetweengasbubblesandsurroundingice.ThestandardmodelisbyHerron&Langway,butitassumestherewasnosnowmelt.Iwillevaluatewhetherthismodelisappropriate.

ArgumentFORsnowmelt(onlymoderatelycertain)

Hobbes:Judgingbythecurrenttemperaturevaluesattheloca+onoftheNEEM-S1core(average-3.34Candstandarddevia+on22.97),theHerron&Langwaymodelmaynotbeappropriateforthiscore.ShouldIusethe<nameothermodelhere>instead?

Step3:Stoplayercoun:ngaDerannualsignalsfadeHobbes:Weareatdepth3525m.IwillcheckwhethertheannualsignalshavefadedtoseeifIshouldkeeplayercoun+ng.

Argument:numberofpeaksperseriesisnormalpartlyapplicablesound

Argument:averagelayerthicknessisnormalmostlyapplicableplausible

Conclusion:con+nuelayercoun+ngmostlyapplicablesound

Hobbes:Wecancon+nuelayercoun+ngajerdepth3525m.

Finaloutput:Reasoningbehindagemodels:•  Performedautomatedlayercoun+ng

betweendepth50mand4020m<thesearemadeup!>

•  UsedDansgaard-Johnsenflowmodelforda+ngdepthsbelow4020mIdon’tseethatabove.Ithinkyou’remissinga4thstep

•  Usedtwodifferentfirndensifica+onmodels:Herron&Langway(orangecurve)and?(bluecurve)

•  Whataboutthegasrecordalignment?•  Etc(moredetailswouldbeincludedinthe

fullreport)

ReproducibilityReproducibilityisakeycomponentofthescien+ficmethod.Ascomputa:onbecomesmorecentraltothescien+ficenterprise,itisurgenttoaddressconcernsregardingreproducibilityofcomputa:onalresults.CSciBoxaddressesreproducibilityconcernsby:•  Tracksallac+onstakenwiththecore,forcomplete

reproducibility•  Includescita+onnotesevery+meyouuseanother

scien+st’stool•  Metadataisalwaysbundledwiththecoredata•  WeuseLinkedPaleoclimateDateformat•  Itiseasy(andmandatory)inCSciBoxtosaveallthe

dataandmetadatathatproducedanagemodel(ifyouwanttosavetheagemodel).Bymetadata,wemeanspecificparameterchoicesandsuch-everythingthat’srequiredtoreplicatethecomputa+onexactly.

•  Itisalsoeasy(butop+onal)tosaveinforma+onabouttheothermodelsthatthescien:stconsideredbutrejected.<Somethingaboutwhythisisimportantforsciencetoo>

•  Hobbes’sknowledgebaseencodesthereasoningbehindgeoscienceresearch.Thisreasoningmaynotalwaysbereportedintradi+onalscien+ficjournaling(?)Theprocessof“algorithmizing”therulesofinferenceforcesscien+ststocodifyandreportalloftheirreasoning.That’scool!

•  CSciBoxspecificallyaddressesmanyoftherecommenda+onsoftheYaleLawSchoolRoundtableonDataandCodeSharing(cite).

Argumenta+vereasoningHobbesconsidersargumentsforandagainsteachtheory,andweighstheirstrength.•  Thismimicsthewayscien:ststhink,andthewaythey

communicateaboutitwithothers—andthusstreamlinescommunica+onbetweenHobbesanditsusers

•  Italsoallowsthescien+sttoconsidermul:pleagemodelsatonce,andenablesHobbestohandleconflic+ngarguments.

•  Atthispointintheprocess,forinstance,thescien+stcanchooseoneortheotherofthesuggestedfirnmodels—ormaychoosetodirectHobbestoexploreboth.

•  Thisistheplacetoputthatmaterialaboutscien+ficdisagreementandconflic+ngrules

Twocomponentsofconfidenceinthereasoning•  Confidenceinthescien:ficknowledge(validity):Isthisruleofinference

valid?Validityrangesfromauniversallyacceptedrule(accepted),toagut-levelfeeling(plausible).

Accepted>Sound>Probable>PlausibleAvaliditylevelisaqachedtoeachrule,andisdeterminedbytheexpertgeoscien:stsontheteam.Thediscretelevelsofvaliditycorrespondtotheweightsscien+stsgivetotheirreasoning,andareeasiertoobtainfromexpertsthancon+nuousvalueswouldbe.

•  Confidenceintheevidence(applicability):Howwelldoesoursitua+onmatchtheruleofinference,asshownbyourdataset?

Highlyapplicable>mostlyapplicable>partlyapplicableHobbesdeterminestheapplicabilityofapieceofevidence,byanalyzingthedata.•  Whenweighingtheevidencefororagainstatheory,Hobbesisdesignedto

intelligentlyconsiderthevalidityandapplicabilityofeachargument..Hobbescombinestheindividualargumentsintoasingleconclusion,describinghislevelofbeliefinthetheory.

I’mstrugglingabitwithwri+ngaboutthetwocomponentsofconfidence.Theconceptisrathersubtle,andhardtoexplainwellinashortblurbonaposter.I’dliketotalkaboutit,butIdon’twanttomen+onitandnotexplainitwell.S+llworkingonthat…Alsonotethatthearrowshouldpointtothesec+on“stoplayercoun+ngajerannualsignalsfade”,butIcan’teasilymakepowerpointdothatrightnow.

Hobbes:Okay[saves\mu=13cminthemetadataandproceeds]

PP11A-2203

[email protected]!

Elizabeth Bradley1, Kathleen Finlinson2, Tyler Jones3, Colin Lindsey3, Brett Israelsen1

Kenneth A. Anderson1, Thomas M. Marchitto3, Laura Rassbach de Vesine1, and James W. C. White3, 1 Department of Computer Science, University of Colorado, Boulder, Colorado, USA

2 Department of Applied Mathematics, University of Colorado, Boulder, Colorado, USA 3 Institute for Alpine and Arctic Research (INSTAAR), University of Colorado, Boulder, Colorado, USA

CSciBox:Anintelligentassistantforda+ngiceandsedimentcores

Themainidea:

CSciBox!

Geoscience

Computerscience

Image&courtesy&of&Dorthe&Dahl2Jensen,&via&Jim&White&Photo:DortheDahl-Jensen

Geoscien+st:•  Hascore•  Wantsagemodel•  Bringsscien+ficexper+se

CSciBox

Output:•  Oneormoreproposedagemodelsforthecore•  Sophis+cateduncertaintyes+mates•  Summaryofthereasoningbehindeachmodel

Hobbes

CSciBox:•  Helpsapplyscien+ficreasoningandalgorithms•  Exploresalterna+vesautoma+cally•  Keepstrackofallthereasoninginvolvedinthemodel

crea+on•  Needtoexplaintherela+onshipbetweenHobbes

andCSciBox

Thisisold;willneeda

newscreenshot

ThismaterialisbaseduponworksponsoredbytheNa+onalScienceFounda+on.Anyopinions,findings,andconclusionsorrecommenda+onsexpressedinthismaterialarethoseoftheauthor(s)anddonotnecessarilyreflecttheviewsoftheNSF.

www.cs.colorado.edu/~lizb/cscience.html!

• Sourcecode(python)availableongithub:github.com/ldevesine/cscience• Butyoudon’thavetoknowpythontorunit;wealsohaveone-clickinstallerstheretoo• GNUpubliclicense;freetomodify/extend/useasyouseefit

Wearetryingtobuild—andsupport—ausercommunity.Pleasejoinus!

CSciBoxisopensource

Intheworks:• Saywhatthefutureworklandscapelookslike• Herearesomesugges+ons:

• MakingStra+Counter&Baconbothwork• HelpingHobbesmakesmartdecisionsaboutwhentorunexpensivecomputa+ons• AddingextensivelytoHobbes’rulebase

Winstrup15cita+on

Cheryl R. Richardson, PhD · POD_Diversity2015 poster.pptx Author: Cheryl Richardson Created Date: 11/16/2015 4:17:44 PM

Click Chemistry A B Crosslinked Particles: N B Fig. 4Final REU Poster.pptx Author David Trebatoski Created Date 20130801133304Z

2019 Signs and Symbols Poster - Extinction Rebellion · 2020. 7. 18. · Title: Microsoft PowerPoint - 2019 Signs and Symbols Poster.pptx Author: p250964 Created Date: 7/29/2019 2:48:27

Oxygen Oxygengreif.geo.berkeley.edu/~driver/Talks/AGU2015.pdf · Introduction:firstprinciples methods for warm dense matter (WDM) First-Principles Equations of State Calculations

Rejuvenation of 7-GeV SuperKEKB Injector Linac · ipac18-furukawa-poster.pptx (読み取り専用) Author: Kazuro Furukawa Created Date: 4/28/2018 10:55:33 AM

Global Solutions for Nonlinear Systems - Computer …lizb/papers/ai-math98.pdf · Global Solutions for Nonlinear Systems Using Qualitativ e ... If a solution exists, ... mo dels of

Maschmeyer AGU2015 compressedseoe.sc.edu/.../Maschmeyer_AGU2015_compressed.pdf · 2016. 2. 8. · Training Data Testing Data Submarine Lava Morphology Typical oceanic ridge lava exhibits

Senior Design Poster.pptx

Palaeontologia Electronica - University of Colorado …lizb/phd/kent.pdfPalaeontologia Electronica Vol. 12, Issue 3; 11A: 26p; VIRTUAL PALAEONTOLOGY: GAIT RECONSTRUCTION OF EXTINCT

X(s) Y(s)lizb/papers/ida-chapter.pdfresp onse h (t) to mo del the system, or in frequency domain, using resp onse transfer function H s). The impulse resp onse of a system is its transien

Measurement of the Half-Life of Orthopositronium...revised positronium poster.pptx Author Mark.Yuly Created Date 3/10/2009 9:33:17 AM

A Cloud Computing Workflow for Scalable Integration of ...dusk.geo.orst.edu/Pickup/Esri/AGU2015/IN52A-03-Urban-Flow.pdf · A Cloud Computing Workflow for Scalable Integration of Remote

Poster.pptx (1)

Johnson Brock poster - Michigan State UniversityTitle: Johnson Brock poster.pptx Author: Martin Brock Created Date: 7/1/2009 12:46:00 PM

LiPD and CSciBox: A case study in why data standards are ...lizb/cscience/agu2016.pdf · • A powerful analysis and design environment for building and working with age models

Recurrence plots of experimental data: To embed or not to ...lizb/papers/1.166372.pdf · points in nonstationary data sets. Trulla et al.,4 for instance, analyzed the dynamics of

E2.05 Senior Design Postere727ba2f-b81e... · Title: Microsoft PowerPoint - E2.05 Senior Design Poster.pptx Author: ryans Created Date: 5/3/2020 2:50:00 PM

Crest Challenge Poster · Title Microsoft PowerPoint - Crest Challenge Poster.pptx Author Jessie Created Date 9/1/2017 9:13:13 PM

Captain America poster.pptx

Immersive Examination of the Qualitative Structure …lizb/qr08/papers/Gruchalla.pdfImmersive Examination of the Qualitative Structure of Biomolecules ... ware and software that provides

Landsat on Amazon Web Services - Dawn Wrightdusk.geo.orst.edu/Pickup/Esri/AGU2015/IN43B-1727-AWS.pdf · Landsat on AWS is the result of collaboration with many AWS customers, including:

Methadone Maintenance Therapy and Viral Suppression …programme.aids2016.org/PAGMaterial/eposters/0_5570.pdfMMT, cocaine, VL suppresssion_AIDS2016 poster.pptx Author: Peter Vann Created

LIII Poster.pptx

High-Resiliency and Auto-Scaling of Large-Scale Cloud Computing …dusk.geo.orst.edu/Pickup/Esri/AGU2015/IN43B-1733-Hook... · 2015-12-15 · Per node transfer Scalable internal data

OceanObservations&from&Below&PetermannGletscher,Greenlandmuenchow.cms.udel.edu/html/talks/AGU2015.pdf · OceanObservations&from&Below&PetermannGletscher,Greenland Andreas&Münchow

satellite - Stanford Universitycs231n.stanford.edu/reports/2017/posters/557.pdfMicrosoft PowerPoint - poster.pptx Author pmaoki Created Date 6/5/2017 12:54:34 PM

IN51C-05 A Look Under the Hood - Dawn Wrightdusk.geo.orst.edu/Pickup/Esri/AGU2015/IN51C-05-Knosp.pdf · 2016-01-18 · ?5+5@5>)"!*4,3+)4+K*)"&2(5+)" ©2015 California Institute of

ACE: Age Calculation Engine—A Design Environment for ...lizb/papers/ADVCOMP07.pdf · sign and construction of cosmogenic dating software, we are creating artificial intelligence

DEVELOPMENT OF MOBILE APP FOR ACCESSING … 2.pdfMicrosoft PowerPoint - islh poster.pptx Author: 55098788 Created Date: 9/21/2016 2:43:46 PM

490B Poster.pptx (2)

NoBODY - part 1 (LizB-UP)

Time-Varying Flow Analysis and Visualization for Climate ...tpeterka/posters/2016/peterka-axiccs16-poster.pdfpeterka-axiccs16-poster.pptx Author: Tom Peterka Created Date: 1/18/2016

Discrete Global Grid Systems A Framework for the Next Era ...dusk.geo.orst.edu/Pickup/Esri/AGU2015/IN43B-1729-Pyxis.pdf · Discrete Global Grid Systems –A Framework for the Next

idj poster.pptx

SAE AERO poster AERO poster.pdfSAE AERO poster.pptx Author: Somsak Sukittanon Created Date: 20120512013335Z