machine learning lab course - tum...machine learning practical course –summer term 18 data mining...

MachineLearningPracticalCourse– SummerTerm18 Data Miningand Analytics

MachineLearningLabCourse

OrganizationalMeeting

SummerTerm 2018 Data Miningand Analytics

lecturer: Prof. Dr. Stephan Günnemann


§ Prof.Dr.StephanGünnemann

§ DanielZügner

Thisisapracticalcourse(Praktikum)forMaster students!Nameofmodule:Large-ScaleMachineLearning(IN2106,IN4192)

website:ml-lab.in.tum.de

Team

2


Why attend our Machine Learninglabcourse?

1. Getthechancetoimplementandapplystate-of-the-artMLalgorithms

2. Gainhands-onexperienceworkingonreal-worlddata,solvingreal-worldtasks(e.g.byworkingononeoftheprojectsbyourindustrypartners).– Successfulprojectsmightevenqualifyforasubsequentmasterthesis.

3. Workonlarge-scaleproblemswiththesupportofstate-of-the-artGPUcomputingresources.

3


§ Requirementsforthelabcourse– strongprogrammingskills(Java,Python,C++,Java,etc.)

– strongknowledgeindatamining/machinelearning

– youshouldhavepassedrelevantcourses(themore,thebetter)

- MiningMassiveDatasets

- MachineLearning

- Ourseminars

– self-motivation

§ Additionalselectioncriteria– otherrelevant experience(projectsincompanies,experienceasaHiWi)

- youcansendanoverviewofyourexperiencetous(seeendofslides)

Requirements

4


§ Groupsof3-4students§ Eachteamwillworkonadifferentproject,e.g.incooperationwithoneof

ourindustrypartnersoronatopictheyhavesuggestedthemselves

§ Groupsareallowed(should)collaborate!– exchangeyourexperiencewiththeothergroups– howdotheothergroupstacklecertainproblems?

§ Technicalaspects:– eachgroupwillgetexclusiveaccesstoatleastonehigh-endGPUserverwith

- 4xNVIDIAGPUw/11GBRAM

- 10-coreCPU

- 256GBRAM

– scaleupyourmodelsanddata!

Organization

5


§ Weeklymeetings(around90-120minutes)– eachgroupshouldbrieflyreporttheirprogress,openproblems,andnextsteps

§ Regulardocumentationofyourwork– statusreportsanddocumentation(wemightsetupawiki)

– useofacentralcoderepository

Organization

6


§ Thegradeis based onthe whole semester‘s performance!– regular completion of documentation

– regular presentations/discussions during semester

– finalpresentationattheendofthesemester

- overviewaboutwhatyouhavedone,howdidyouimplementit,whataretheresults,whatwentwrong,discussionoftheframework,…

- eachmemberoftheteamneedstopresentsomeparts

Grading

7


§ Techniqueswemightwanttolookat(ifyouknowthese,that'sgood!)– Optimization(e.g.viagradients)

– Stochastic optimization

– Neural networks

– Learningwithnon-i.i.d.data(e.g.temporaldata)

§ Tasks:– preprocessing

– classification

– profiling

– clustering/topicmining

– recommendation

– anomalydetection

– …

Content

8


Projects

There are three types of projects inthis labcourse:

Academicprojects

Industryprojects

Your ownprojects

9


Reproduction and improvement of apublished model

§ Canyou spot inconsistencies inarecent publication‘s experimentalsetup?Canyou even improve their results?

§ Studentscanchoosearecentalgorithm(e.g.fromICLR2018),andaimtoreproduceandimprovetheresultsinthepaper.

§ Giventhecomputationalresourcesavailabletothestudents,theycanevenselectlarge-scalemodelsandevaluatethevalidityoftheresultsandclaims.

§ Thiscanalsobeagoodwaytolaythefoundationofanewalgorithmforamasterthesis.

10


Industry project:Oktoberfestfood classification

§ Industry partner:ilass AG,maker of software for gastronomy and partytents (e.g.Oktoberfest).

§ Theproject willbe about detecting and classifying food items onimages tobe extracted from avideo stream.

§ Representativepresenttoday:PeterVogel

11


Industry project:Automatic anonymization of faces

§ Automatic anonymization of faces inimage and video data is important toprotect the privacy of people.

§ Blurring or completely graying outparts inimages where faces aredetected means aloss of information since allfacial features are removed.

§ Goal:develop amethod for face anonymization while preserving themostrelevantfacial features to stillrecognize basic information likeemotions.

12


Industry project:Siemens

§ Detailsto be announced.

13


Own projects

§ You can submit abrief exposé of your project idea provided that:– There is aconsiderable challenge from amachine learning perspective,e.g.

non-i.i.d.data (graphs,temporaldata),very noisy data,new application,

– You have asufficiently largeand challenging dataset athand (e.g.from anopendata platform),

– Theproject is suitable for agroup of 3-4students.

14


Own projects:exposé

§ Theexposé should contain– abrief description of the problem and why it is important,

– adescription of the dataset you planto use

– arough outline of anapproach you would liketo pursue

§ If you are agroup of students,only one student should fill inthe exposéand add the others‘student ID

§ Max,3,000characters

§ Submit viaonlineform(see endof slides)

15


Registrationviathematchingsystem!

Modulename:Large-ScaleMachineLearning(IN2106,IN4192)

+fillouttheapplicationform(seenextslide)

Registration

16


§ Filloutourbriefonlineformaboutyourexperienceuntil14.02.2018– youcanprovideuswithalistofyourexperienceindatamining/machine

learning(courses,projects,…)

– pleasesendashortoverviewonly(bulletlist);notacompleteCV

– (optional)attachabrief exposé of your own project idea.

§ Checkml-lab.in.tum.de for alinkto the form.

YourExperience

17

machine learning lab course - tum...machine learning practical course –summer term 18 data mining...

Documents