E vo Arm Custo miza tio n! Dime nsio ns De sig n c o nstra ints • Sma ll ro b o t a rm • E ve ry 3D-printe d a rm c a n Se rvo c o nstra ints b e diffe re nt • 3 de g re e s o f fre e do m • Cha ng e me c ha nic s fo r • 3D printa b le diffe re nt purpo se s • Co ntro lle d with a Pytho n RoboVis App Co nfig I K So lutio n • … whe re to g o fro m he re ? Alista ir Wic k E xplor ing Possible Configs Dime nsio ns Dime nsio ns De sig n c o nstra ints De sig n c o nstra ints Se rvo c o nstra ints Se rvo c o nstra ints Vis c a n he lp! Cha ng ing o ne o f do ze ns o f pa ra me te rs Co nfig I K So lutio n Co nfig I K So lutio n T e dio us a nd impra c tic a l to try ma ny de sig ns Diffe re nt pe o ple ne e d diffe re nt c a pa b ilitie s Ca n the e xplo ra tio n pro c e ss b e ma de a c c e ssib le ? Wa nt to ra pidly ite ra te – ne w/ diffe re nt c o nfig s Vis Ide a WHAT PITCH: I nte ra c tive e xplo ra tio n o f de sig n spa c e VISUALIZING THE ENERGY • ENERGY PERFORMANCE DATA OF A BUILDING (FOR NOW THE BUILDING IS THE CENTER FOR INTERACTIVE RESEARCH ON SUSTAINABILITY/”CIRS”) Data: Ca lc ula te d o nline PERFORMANCE OF A BUILDING • TIME-SERIES DATA FROM SENSORS INCLUDING TEMPERATURE AND OCCUPANCY DATA (IF Re a c ha b le po ints POSSIBLE) Ma x lo a d (a c ro ss re a c ha b le spa c e ) ARASH SHADKAM • DERIVED: NORMALIZED ENERGY PERFORMANCE DATA Ma x ve lo c ity (a c ro ss re a c ha b le spa c e ) L e ng th 0 L e ng th 0 De sign: L e ng th 1 L e ng th 1 Spa tia l da ta -> spa tia l displa y? De rive a ttrib ute s? Co mb ine c e rta in pa ra me te rs? 120N 110N 100N 90N 80N L e ng th 0 L e ng th 1 HOW HOW WHY • FACET: MULTI-FORM OVERVIEW-DETAIL VIEWS/LINKED HIGHLIGHTING • BETTER UNDERSTANDING OF THE BUILDING’S ENERGY PERFORMANCE • MANIPULATE: SELECT • DISCOVERING TRENDS AND CORRELATIONS IN THE ENERGY PERFORMANCE DATA AND • REDUCE: FILTER/RANGE SLIDERS FOR DIFFERENT TIME SPANS IDENTIFY POTENTIAL OPTIMIZATION OPPORTUNITIES IN THE BUILDING’S PERFORMA NCE THANKS!
WHAT IS COMPUTER PROGRAM PERFORMANCE HOW DO VISUALIZATION TOOLS HELP? PROJECT OBJECTIVES DEBUGGING? A VISUALIZATION TOOL FOR Let’s look at an existing visualization tool… Diagnosing why a computer program is running slowly Create a visualization tool which: COMPUTER PROGRAM PERFORMANCE DEBUGGING - Uses the “search , show context , expand on demand” Augustine Wong approach - Visualizes “patterns” of computer program behavior - Evaluates which patterns are good starting points for initially exploring the computer program 2 3 4 Mo(va(on Mo(va(on Quantum Annealing Visualization • Strengths of ML allowed expansion to diverse fields • Biggest factor for users is understanding how Visualiza(ons For Jus(fying Austin Wallace predic<ons occur 5th year undergraduate student Integrated Science-Machine Learning Machine Learning Predic(ons • Fields and contexts far removed from tradi<onal ML Chimera Graph • Par<cularly important in 1 : David Johnson • High risk applica<ons like medicine • Users not trained in ML • Consumer-facing applica<ons such as Recommender Systems • Context-Aware applica<ons 1 Biran, McKeown. . 2014. Jus<fica<on Narra<ves For Individual Classifica<on Jus(fica(on Visualiza(on Motivation The Dataset • Visualiza<ons present important evidence for a https://www.yelp.com/dataset_challenge/dataset predic<on Yelp - Target User: Yelp end-users Visualization Tool - Comparing businesses - Filtered visualization • Intensions are to <e in to thesis work Dilan Ustek Matthew Chun Munich Scope Project ¡Pitch One city but yet to be decided ● Focus on the end users, aka the people who use the Yelp site/app ● Data features to consider ... it depends but theme of holistic/detailed ● comparison Information ¡Visualization 2017 ○ Discover the “nuances” behind the existing Yelp data eg. distribution of 5 star restaurants in different price categories Felix ¡Grund More informed decisions for end users ○
Who ¡is ¡Scandio? What ¡is ¡a ¡fixed ¡price ¡project ¡at ¡Scandio? What ¡are ¡the ¡project ¡results? What ¡are ¡the ¡key ¡attributes? • 2016: • Efforts ¡range ¡from ¡5 ¡days ¡-‑ 100 ¡days • Total ¡amount ¡of ¡efforts ¡in ¡the ¡end 1. Hours ¡worked – 40 ¡employees • Duration ¡ranges ¡from ¡3 ¡weeks ¡– 1 ¡year – Exactly ¡as ¡estimated ¡(rare) – Employees ¡track ¡time ¡on ¡project ¡in ¡web ¡app 2. Degree ¡of ¡completion ¡(DOC) – 82 ¡clients – Less ¡than ¡estimated ¡(sometimes) ¡ J • Before ¡project ¡starts: ¡effort ¡estimation – 176 ¡projects – More ¡than ¡estimated ¡(sometimes) ¡ L – Estimated ¡monthly ¡by ¡project ¡lead • Generally ¡higher ¡risk ¡of ¡“failure” 3. Hourly ¡rate ¡for ¡project • Projects: – If ¡over ¡estimation ¡in ¡the ¡end, ¡company ¡mostly ¡has ¡ – Determined ¡in ¡the ¡beginning ¡dependent ¡on ¡ – Fixed ¡price (“client ¡pays ¡what’s ¡estimated”) to ¡pay ¡(sometimes ¡compromises ¡with ¡client) budget ¡and ¡total ¡effort – Time ¡and ¡material ¡(“client ¡pays ¡the ¡hours”) – Changes ¡retrospectively ¡depending ¡on ¡1 ¡and ¡2 ? ¡Questions ¡? Time ¡Tracking Project ¡results ¡(good) • When ¡do ¡estimation ¡and ¡degree ¡of ¡completion ¡conflict? • When ¡are ¡our ¡hourly ¡rates ¡too ¡low? • How ¡do ¡hourly ¡rates ¡change ¡retrospectively? Is ¡there ¡still ¡time? • What ¡tendencies ¡can ¡we ¡observe ¡over ¡multiple ¡projects? • When ¡interfere ¡to ¡maintain ¡project ¡success? • How ¡can ¡we ¡identify ¡wrong ¡estimations ¡on ¡DOC? • How ¡do ¡project ¡leads ¡differ ¡in ¡their ¡monthly ¡estimations? • … Project ¡results ¡(bad) Background: Visualizing Internal What is Machine Learning Components of a Machine Learning is taking over. Thanks. Convolutional Neural Applied to many fields: Bioinformatics, Gaming, Medical diagnosis, Marketing, Machine Vision, …. Network Mahdi Ghodsi - Hooman Shariati Convolutional Neural Network Convolutional Neural Network Convolutional Neural Network Convolutional Neural Network Very Popular Research Area The idea has been around since 1980s But Introduction of GPU computing with 30x speed up gave DNNs a boost However ... ImageNet Competition Google Deep Dream
Convolutional Neural Network Convolutional Neural Network Visualizing and making sense of of CNNs in literature: Visualizing and Understanding Convolutional Networks By M. Zeiler (NYU) How researchers see CNNs How researchers see CNNs How CNNs looks like “Neural networks have long been known as “black boxes” be- cause it is difficult to understand exactly how any particu- lar, trained neural network functions due to the large number of interacting, non-linear parts.” Yajin Zhou Department of Computer Science North Carolina State University Tamoxifen 10-year risk esGmates Case Scenario Point esGmates … compared to 5-year risk esGmates (out of 1000) • Imagine you are • Imagine Be;y only cared about her chance of Be;y A"ribute Change dying from breast cancer and her chance of Breast cancer recurrence ê 28 Visualizing Ambiguity • Just finished developing endometrial cancer chemo for Death from breast cancer ê 28 5-year vs. 10-year Tamoxifen Therapy breast cancer 31 James Hicklin • Typical post- Number of People (out of 1000) Development of gallstones é 2 26 chemo therapy 21 is Tamoxifen for 16 Development of endometrial é 16 11 5 years cancer 6 Stroke é 2 1 Decrease in deaths from BC Developing Endometrial Cancer With confidence intervals … AlternaGves to Error Bars Project Violin Plots Box Plots • Design new visualizaGon to present ambiguity 5-year vs. 10-year Tamoxifen Therapy 36 to paGents 31 Dviz • InteracGvity Number of People (out of 1000) 26 – Adjust bounds of error 21 h;p://www.datavizcatalogue.com/methods/violin_plot.html h;p://www.datavizcatalogue.com/methods/box_plot.html – Show best & worst case scenarios 16 Visualizing Distributed Systems – Show how risk esGmates might change given Dynamic Icon Arrays with Stewart Grant and Jodi Spacek Gradient Plots 11 different samples 6 1 Decrease in deaths from BC Developing Endometrial Cancer Motivation Concept etcd (distributed key value store) puts -> gets Limitations ● Understanding the behaviour of distributed systems is hard ● Collect distributed snapshots (state from across the whole system) ● States are not labeled meaningfully ● Developers need tools for comprehending their systems ● Calculate a distance between each snapshot (xor distance) ● Semantics of state transitions are not clear ● Most distributed systems are designed around FSM ● Plot each snapshot at it’s relative distance using clustering ● FSM’s require both FSM are often how developers think of their systems Connect each snapshot with a time curve ● ● ● Can an FSM be generated from an execution so developers can check their mental models?
Recommend
More recommend
