Network Learning: AI-driven Connectivist Framework for E-Learning 3.0
2 likes841 views
This document discusses the evolution of eLearning and introduces a connectivist framework for eLearning 3.0. It summarizes eLearning 1.0 which focused on reading content and behaviorism/cognitivism theories. eLearning 2.0 allowed writing and social interaction and incorporated constructivism and social learning theories. However, most created content is unused, redundant, or results in information overload. The document proposes connectivism which views knowledge as distributed across networks and learning as constructing/navigating these networks. It introduces a conceptual framework using AI to connect content, people, and models through different layers and modules.
![Connecting/Pruning Nodes
Network Structure Learning
Proposed Method Proposed
Proposed Method Propo
Learning in Collaborative Settings Network Structure Learning yt+1 yt+1 yt+1
yt yt yt
input index
input index
input index
input index
input index
input index
input index
input index
Model
0.4
a)a) Adding training point causes many b)
Adding training point causes many b) Ad
output estimates toto change.
output estimates change. out
output feature index output feature index output feature index
0.35
SL outpu
Traditional Collaborative Proposed Method Proposed Approach
Proposed Approach
Proposed Method
0.3
0.25
Proposed Approach
Proposed Approach
yt+1 2
Use ychanges in in the estimates ⇥yt yt+1 ⇥2⇥to
P (yt+1 )
Learning in Black-box Settings 0.2 Use changes the estimates ⇥yt
Figure 5: T1 =⇤yt yt+1 ⇤ and the yt+1 yt+1 y t+1
generalization error G d )
generalization error G (x(xd )
t+1
0.4
0.4 value that it tries to approximate ⇥G
0.15
yt y y
0.35
(Section Section 3.1). Most importantly, yt t
t
0.1
0.35
0.3 high values of ⇤yt yt+1 ⇤2 should cor-
input index
input index
where yt+1 are the estimates after d , y, ) )
where yt+1 are the estimates after (x(xd d ydw
input index
input index
16% 0.3
84% respond to high values of ⇥G, since
those are the points that are likely to be
0.25
0.25 P (yt+1 )
chosen.
0.05
P (yt+1 ) 0.2
0.2
Figure 5: T =⇤y y ⇤ and the 1 t t+1
0.15
Figure 5: that it tries to t+1 ⇤ and the ⇥G
value T1 =⇤yt y approximate
0
0.15
value(Section tries to 3.1). Most importantly,
that it Section approximate ⇥G
y y
high values of ⇤yt yt+1 ⇤2 should
(Section Section 3.1). Most importantly, cor-
9
0.1
16% 84% high values of to high t+1 ⇤2 should cor-since
respond ⇤yt y values of ⇥G, Proposed
0.1 yt+1 A−optimal
16% 84% respond toare thevalues of ⇥G, likely to be
those high points that are since D−optimal
thosechosen. 8points that are likely to be
are the
0.05 E−optimal
Transductive
chosen.
0.05
Random
Optimal
Figure 3: Distribution of the ⇥ estimates
0
y y
⇥ 0 9 7
Proposed
yt+1 in relation to the estimate yt and the
y y
y t+1 9
A−optimal
D−optimal
Proposed
E−optimal
Mean Squared Error
y t+1 8 A−optimal
Transductive
f (x) f (x)
true value y.xFigure 3: Distribution of y estimates
D−optimal
Random
6 E−optimal
x
8 Optimal
yx the
Transductive
x Random
yx = yx =
7
·x ⇥ x
Figuret+1 in relation to·the estimate yt and the
y 3: Distribution of the estimates 7
Optimal
Mean Squared Error
true value y.
ˆ yt+1 in relation to the estimate yt and the
6 5
a)a) Adding training point causes many b) Adding training point causes few
Mean Squared Error
yt+1 true value y.
(a) White (x) Model.
f Box (b) Black Box Model.
Adding training point causes many b) Adding training point causes few
6
x
5
yx ˆ
yt ˆ
y y*
yx = ·x
t+1 4
5
ˆ
yt+1
output estimates to change. output estimates toto change.
output estimates change.
ˆ
yt y* c
Figure 4: For Black Box models only the inputs and outputs
a b 4
output estimates to change.
a b * c
ˆ
y y t 4
(b) Black Box Model.
c 3
a b 3
are accessible, internal workings areFigure after4:after after training unlike for
Figure 4: not accessible, point
yFigure the training point
y the training
is added to the training set point
4: y the (making the
3
is added toadded to the training set (making the the
is the training set (making
2
1 2 3 4 5 6 7 8 9 10
models only themodels [11].outputs of trainingofpoints points equal to t+1).
White Box inputs and number
Training Set Size
2
2 1 2 6 3 4 5 6 7 8 9 10
1 2 3 4 5 7 8 9 10
number number of training pointsequal to t+1).
equal to t+1).
Proposed Approach
Training Set Size Training Set Size
kings are not accessible, unlike for Figure 6: Evaluation of active learning
Figure 6: Evaluation of active Error: lower
criterions (Mean Square learning
criterions (Mean Square Evaluation at active learning
values are better, values are different of
Proposed Approach
yt+1 yt+1 Figure 6: Error: lower
yt+1 2 estimate improvement in the
the statistical significance level of 95%).
values are better, values are different at
criterions (Mean Square Error: lower
Use changes in in the estimates ⇥yt yt+1 ⇥2⇥toto estimate improvement in the
Use changes the estimates ⇥yt
yt yt
input index input index the statistical significance level of 95%).
values are better, values are different at
a) Adding a training point
yt+1
yt
b) Adding a training point the statistical significance level of 95%). generalization error G d )
generalization error G (x(xd )
influences many output es- input index
influences only a few output](https://image.slidesharecdn.com/2012-03-jst-rubens-slideshare-120320213309-phpapp01/85/Network-Learning-AI-driven-Connectivist-Framework-for-E-Learning-3-0-22-320.jpg)
Network Learning: AI-driven Connectivist Framework for E-Learning 3.0
- 1. 国立大学法人 電気通信大学 振興調整費 Unique and Exciting Campus ネットワーク ラーニング Network Learning AI-driven Connectivist Framework for E-Learning 3.0 Neil Rubens Active Intelligence Group Knowledge Systems Laboratory University of Electro-Communications Tokyo, Japan
- 2. Evolution of eLearning: eLearning 1.0 eLearning uses technology to enhance Learning --------- ‣ eLearning 1.0: --------- --------- --------- --------- ‣ reading: content became easily accessible ‣ logging: user’s activities could be logged and analyzed ‣ Learning Theories: ‣ Behaviorism: learning is manifested by a change in behavior, environment shapes behavior, contiguity ‣ Cognitivism: how human memory works to promote learning
- 3. Evolution of eLearning: eLearning 2.0 ‣ eLearning 2.0: --------- --------- --------- --------- --------- --------- --------- --------- --------- --------- ‣ writing: anybody can easily create content (e.g. blogs, wiki, etc.) ‣ socializing: interaction is easy (e.g. facebook, twitter, etc.) ‣ Learning Theories: ‣ Constructivism: constructing one's own knowledge from one's own experiences (enabled through writing) ‣ Social Learning: people learn from one another (enabled through socializing)
- 6. Broken Knowledge Cycle ‣ Problem: The current cycle of knowledge creation/utilization is inefficient ! ‣ large portion of created content is never utilized by others* only 0.05% of twitter messages attracts attention (Wu et. al., 2011) only 3% of users look beyond top 3 search results (Infolosopher, 2011) ‣ large parts of created contents are redundant (Drost, 2011) ‣ Peak Social – the point at which we can gain no new advantage from social activity (Siemens 2011) utilize U0lized dge no wle K is ting Ex Redundant create Knowledge Novel *there are some personal benefits e.g. externalization, crystallization, etc.
- 8. Information Overload for Computers not a Problem but an Opportunity Tan, Steinbach, Kumar; 2004 500 Points 2,000 points 8,000 points
- 9. Messaging Networks How can we use computers to learn in these settings? http://datamining.typepad.com/photos/uncategorized/2007/04/08/twitter20070405.png Citation Network Social Network hp://wiki.ubc.ca/images/f/ff/SocialWeb.jpg9 hp://www.kieranhealy.org/files/misc/SocCoreCites.jpg:
- 10. Our Focus informa0on connec0vity Nova Spivack© social connec0vity
- 11. Connectivism (Learning Theory) Connec0vism: Knowledge is distributed across a network of connecTons, and therefore learning consists of the ability to construct and traverse these networks (Siemens Downes, 2008) Property Behaviourism Cognitivism Constructivism Humanism Connectivism Learning Thorndike, Pavlov, Watson, Koffka, Kohler, Lewin, Piaget, Maslow, Siemens, Downes theorists Guthrie, Hull, Tolman, Skinner Piaget, Ausubel, Vygotsky Rogers Bruner, Gagne How learning Black box—observable Structured, Social, meaning Reflection on Distributed within a network, social, occurs behaviour main focus computational created by each personal technologically enhanced, learner experience recognizing and interpreting (personal) patterns Influencing Nature of reward, punishment, Existing schema, Engagement, Motivation, Diversity of network, strength of factors stimuli previous experiences participation, experiences, ties, context of occurrence social, cultural relationships Role of Memory is the hardwiring of Encoding, storage, Prior knowledge Holds changing Adaptive patterns, representative memory repeated experiences—where retrieval remixed to concept of self of current state, existing in reward and punishment are current context networks most influential How transfer Stimulus, response Duplicating knowledge Socialization Facilitation, Connecting to (adding) nodes and occurs constructs of “knower” openness growing the network (social/ conceptual/biological) Types of Task-based learning Reasoning, clear Social, vague Self-directed Complex learning, rapid changing learning best objectives, problem (“ill defined”) core, diverse knowledge sources explained solving
- 12. ConnecTvism: Nice Theory Need: Tools Frameworks To make it Prac0cal hFp://imgs.sfgate.com/c/pictures/2011/12/19/ba-‐BRIDGE20_SFC0105724887.jpg
- 13. Methods
- 14. Conceptual Framework network Analysis Layer --------- --------- --------- --------- connections --------- AI --------- --------- --------- --------- --------- --------- --------- --------- --------- --------- Aggregation Layer links use AI to: Linking Layer § connect contents § connect people nodes § connect people contents § connect models Extraction Layer documents
- 15. Modules Concept concept concept Extraction concept concept concept concept concept concept concepts Group Formation documents tasks concept concept context (documents) Semantic concept concept Influence Mapping Estimation concept concept concept concept concept concept concept concept concept concept interaction log concepts network Analysis Layer concept concept connections Aggregation Layer Knowledge Level concept concept Estimation links concept concept concept concept concept concept concept concept concept concept Linking Layer nodes concepts Extraction Layer documents
- 16. Analysis Layer Link Layer Module Pipeline docs want to learn about: concept Search Engine (Example) Concept concept concept docs Extractor concept concept concept concept concept concept concept concept concept concept Semantic concept concept concept concept concept concept Mapping concept concept concept concept concept concept concept concept Search Engine concept concept concept concept … discussions … concept … concept … concept … concept concept
- 17. Sequence Diagram (Example) user system user system user system I want to know about term t_i I think t_i and t_k are similar .. I want to know about term t_i and t_k social semantics semantics t_i t_i u_i: I think t_i is same as t_j … t_i t_i u_j: no t_is is more like t_p ... u_j: no t_is is more like t_p ... t_i t_i t_i t_i u_m: I think t_i and t_k are similar .. u_i: you are right t_i t_i t_i t_i t_i semantics t_i t_i contents contents t_i t_i t_i t_i contents t_i social social social u_i: I think t_i is same as t_j … u_i: I think t_i is same as t_j … u_j: no t_is is more like t_p ... u_j: no t_is is more like t_p ... u_k: you are both wrong t_i is ... u_j: no t_is is more like t_p ... u_i: I think t_i is same as t_j … u_j: no t_is is more like t_p ... u_j: no t_is is more like t_p ... u_m: I think t_i and t_k are similar .. u_i: you are right
- 18. Connecting Representations ⇤ V (d1 ) Semantic + Network ⇤ V (d2 ) ⇤ Type! Network (N)! Content (C)! Proposed Hybrid! V (d1 ) ⇤ V (d3 ) ⇤ V (d2 ) Characteristics! This is a Title of a ! tfi,j = P ni,j Research Paper ⇤ (d V (d )1 ) V k nk,j Joe Fakeman Jane Noman ! 3 Analysis! V (d1 ) Nowhere University {fakeman, noman}@nowhereuni.edu |D| Abstract idfi = log ~~~~~~~~~~~~~~~~~~~~~ ni,j |{d : ti ⌅ d}| tfi,j = P !#$%$'()%*+(,( !#$%$'()%*+(-( ~~~~~~~~~~~~~~~~~~~~~ ! ~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~ k nk,j Introduction ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~ V (d2 ) tf-idf i,j = tfi,j ⇥ idfi same as C part of N ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ |D| ⇤ ! V (d1 ) ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ idfi = log ! ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ V (d2 ) ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ |{d : ti ⌅ d}| w (d1 ) V = tf-idf i,j i,j User Interface! ! ⇤ V (d2 ) ./+$(/.(,( V (d3 ) This is a Title of a tf-idf i,j = tfi,j ⇥ idfi 2 3 Research Paper w1,j 6 w2,j 7 Joe Fakeman Jane Noman ! ⇤ ! 6 7 V (d3 ) V (dj ) = 6 . 7 Nowhere University wi,j = tf-idf i,j V (d ) 4 . 5 ⇤ {fakeman, noman}@nowhereuni.edu Abstract . V (d3 ) ~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~ 2 wt,j ~~~~~~~~~~~~~~~~~~~~~ 2 3 Introduction ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ 6 w1,j 7 ni,j Deployment! ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ⇤ 6 w2,j 7 tfi,j = P ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ V (dj ) = 6 . 7 4 . 5 ./+$(/.(,( ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ . ! k nk,j ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ wt,j V (d3 ) This is a Title of a ⇤ ⇤ |D| Research Paper V (di ) · V (dj ) idfi = log sim(di , dj ) = ⇤ ⇤ |{d : ti ⌅ d}| Joe Fakeman Jane Noman Nowhere University {fakeman, noman}@nowhereuni.edu V (di ) V (dj ) Execution Abstract ~~~~~~~~~~~~~~~~~~~~~ ⇤ Speed! ./+$(/.(,( V (d1 ) tf-idf i,j = tfi,j ⇥ idfi ~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~~~~~~ Introduction ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ⇤ V (d2 ) wi,j = tf-idf i,j ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~~~~~~~- ~~~~~~~~~~~~~~~~~~ Reverse could also be done: i.e. converting textual representation to a network one. ⇤ 2 3 term-space V (d3 ) 6 w1,j 7 1 ⇤ 6 w2,j 7 V (dj ) = 6 . 7 . tfi,j = n P i,j 4 . 5 ⇥ ! k nk,j w|T |,j w1,j V (d1 ) ⇧ ⌃ 1 ! ⇤ ⇤ ⇧ w2,j ⌃ V (d1 ) |D| idfi = log sim(di , dj ) = V (di ) · V (dj ) ⇧ . . ⌃ ! |{d : ti ⌅ d}| ⇤ ⇤ V (di ) V (dj ) ⇧ . ⌃ V (d1 ) ! ⇧ ⌃ ! — V (d2 ) ⇧ w|T |,j ⌃ V (d1 ) —- ! Network Content based ⇥ idfi tf-idf i,j = tfi,j ⇥ ⇧ ⌃ V (d2 ) V (dj ) = ⇧ ⇧ ⌃ ⌃ ! wi,j i,j = (i, j) j) w = dist(i, ⇧ w1,j ⌃ V (d2 ) != wV (d tf-idf i,j ⇧ ⌃ ! —- i,j 3 ) 2 3 ⇧ w2,j ⌃ V (d2 ) Generalized Assignment Problem ! w1,j ⇧ ⌃ V (d3 ) 2 Objective: given a paper p, 1,j ⇤ 3 6 choose a 6 w2,jof 7 group 7 experts ⇧ . . ⌃ ! ! V (d1 ) w V (d ) = 6 M ⇥ M (of a fixed size6s) 2,j 7jcollectively (d ) ⇤ w that 4 ! possesses . . V 5 . 7 ⇤ . ⌅ V (d3 ) V (d ) = 6 6 7 1 ! the most expertise about p: . 7 j 4 . 5. w|N |,j w|N |,j V (d3 ) wt,j P maximize R(M, p) = m2M r (m, p) (1) ! subject to |M | =!(d ) s (2) V (d2 ) ⇤ ⇤ V (di ) · V (dj ) V 2 Challenge sim(di , dj ) = ⇤ ⇤ V (di ) V (dj ) Expertise is not additive: 1 ! X ! V (d3 ) R(M, p) ⇤= V (d3 . r (m, p)) (3) m2M – Group Expertise Estimation Assumptions Mp are athors of paper p; so R(Mp , p) = 1, and ⇤ ⇤ network-space R(M, p) = 0, where M ⌅ Mp = Ø. ⇤ term-network-space M ⌅ M⇤
- 19. #pouf #noel #deco #mobilier #facebook Connecting Concepts #sustainable #evenementiel #savings #twitter #cadeau #reduce #buzz #design #reuse #money Modeling of: #inception #fuelcell #web #prop23 #hydrogen #recycle #cleantech #mobile #diy #new #florida • topics #pr #innovation #california #news #global #home #biofuels #biofuel #greenroofs #publicrelations #boerse #asia #blog #thermal #frugal #biz #windenergy #aktien #hot #business #greenit • lexicon #gossip #improvements #budget #cleanthinking #nw #change #construction #lightingdesign #warming #socent #benefits #finland #tudelft #eco #ff #green #audit #finance #ca #aia #doe #electric #cofely #scriptie • semantics #duurzaam #architecture #health #india #efficient #greenbuilding #ecomonday #stock #footprint #renewableenergy #energy #tips #tip #carbon #ev #ecofriendly #ac #cancun • dynamics #technology #homes #israel #china #greentip #climate #energyefficiency #delonghi #pv #solar #canada #co2 #csr #building #economist #alternative #education #btu #45 #dd45p #pt #dd50p #50 #science #re #cleanenergy #clean #electronics #cre #10000 #lg #art #africa #coal #dehumidifiers #novosti #climatechange #conservation #electricity #cooperative #emissions #edgestar #star #frigidaire #animals #nasa #environment #cop16 #efficiency #cville #datacenter #cdw #dehumidifier #rated #earth #dlr #pint #space #nature #sustainability #windpower #cars #65 #microwave #energysaving #it #70 #b #yahoo #sandiego #economy #soleus #c #renewables #greenenergy #oil #sd #peakoil #ontario #startup #dp1 #a #un #d #neweconomy #wef #encell #vc #dimmer #air #agw #water #gobeyondoil #wcs #pakistan #seattle #ecosmart #30 #sunpentown #globalwarming #unfccc #energyrevolution #iea #solarpower #ft #investment #us #uk #renewable #pkfloods #eecbg #platts #lamp #hybrid #011013 #epa #transport #solpwr10 #pentair #intelliflo #climategate #socialmedia #energystar #wind #products #led #vs #enviro #climaterealism #turbines #living #power #car #leaf #offshore #housing #sun #greenjobs #nissan #p2 #politics #cnet #google #career #meter #sunpower #utility #lighting #tech #london #greentech #tv #europe #dems #apple #organic #mass #topprog #gadgets #travel #greenliving #nuclear #eu #becktips #gadget #greenbuild #smartgrid #texas #dhilipsiva #baby #leed #cisco #rural #free #gop #tcot #obama #technews #tfb #taf #tcn #cloud #media #gridweek #gas #victoria #3 #fail #pledge #livewithless #glam #privacy #save #lx #streetlights #ftrs #p21 #careers #saving #r #hhrs #youcut #thepowerofwaiting #policy #leds #rs #winterhomeprep #football #veterans #smartmeter #ledstreetlights #ocra #lampe #windworks #ghg #sp1580 #momentive #motiongraphics #tlot #sgp #timelyadvice #win #underoath #licht #clothdiapers #arm #jobs #ebc #utilities #fuel #environmental #siemens #light #glennbeck #tpp #disambiguation #affiliate #linux #employment #hayward #2 #9 #twisters #reddit #hiring #46201 #912 #geospatial #marketing #fossil #teaparty #rulez #engineering #silver #6 #4 #heatpumps #heathrow #1 #capandtrade #gis #saveenergy #uksnow #job #digguser #fuels #realestate #pump #jobdanmark #in #digg #gold #sale #sp2610x15 #5 #8 #ontpoli #hohoho #mlfeeds #cdnprog #trhug #super #7 #stopglobalwarming #starbucks #architekt #cdnpoli #ondp #tweetmyjobs #calau #vt #cookies #aps #99 #indee #getinthegame #photography #indeed #credits #haier #pod #sfo #tax #esncm053e #management #jobangels #solarthermal #grid #meters #breakingnews #accenture #fb #smart #wpccc #dms #cochabamba #dsm #stocks #usa #cmpcc #noc #bolivia #cfd #mkt #webhosting #webhost #greenwebhosting Figure 7: Visualization (zoomable) of 416 Hashtags in Energy-Related Tweets, September 2010 Through January 2011; created with Network Explorer (Rubens et al., 2011). proaches that help understand the large scale conversations enhanced word-of-mouth to stimulate user-generated per-
- 20. epilepsy differentiation Domain-specific therapy surgery migration atrial fibrillation glaucoma proliferation review mri smoking cytotoxicity p53 blood pressure Semantics echocardiography reactive oxygen species nutrition mitochondria hypertension gender apoptosis antioxidant diet oxidative stress insulin growth mutation diabetes physical activity infection metabolic syndrome 2 diabetes type metastasis sepsis insulin resistance radiotherapy complications phylogeny cytokines exercise refrontal cortex chemotherapy evolution stress body mass index inflammation lung cancer eeg fmri gene expression microarray anxiety cancer obesity metabolism depression atherosclerosis drinking water bangladesh ttention serotonin quality of life chronic obstructive pulmonary disease arsenic dopamine cardiovascular disease perspective multiple sclerosis schizophrenia breast cancer copd education outcome ethics laparoscopy immunohistochemistry internet fish adherence communication adolescents mercury pediatric prognosis asthma heart failure particulate matter lead trauma survival fertility myocardial infarction cadmium air pollution environmental health environmental exposure antibodies mortality pregnancy health korea children child occupational exposure climate change polymorphism adaptation polychlorinated biphenyls environment pesticides public health genetics prevention biomarkers exposure epidemiology biomonitoring incidence india hivrisk factors prevalence transmission dispatch outbreak brazil human africa zoonoses tuberculosis research surveillance china antimicrobial resistance letter viruses pneumonia management diagnosis molecular epidemiology bacteria treatment west nile virus drug resistance pcr osteoporosis parasites influenza elderly vaccine genotype malaria vaccination disease outbreaks
- 21. P95-1032 P91-1022 W96-0201 Connection Formation J93-1006 W93-0301 P96-1018 P93-1001 P00-1050 P97-1039 P97-1038 P93-1002 P91-1023 P94-1012 A97-1050 C98-1066 A94-1006 J97-2004 J96-1001 Curriculum / Survey Design C94-2178 C98-2129 P97-1063 P93-1003 H91-1026 P93-1004 J93-1004 Original Design: W95-0115 P06-1062 J00-2004 C98-2153 W01-0505 P95-1033 W03-0301 H05-1011 Lacks Information P05-1057 J03-3002 W02-1012 C96-2141 P97-1037 W05-0812 J05-4003 P06-2014 P06-1065 P07-1001 J93-1003 P04-1066 H05-1010 N07-2022 C00-2163 P06-1097 D07-1006 W08-0409 N06-1014 J07-3002 P03-1012 P06-1009 H05-1012 N01-1026 C04-1006 P06-1002 J90-2002 J93-2003 C04-1032 N06-1013 H05-1022 P07-1039 Limited Diversity P03-1041 H05-1009 D09-1050 E06-1019 P07-1003 P08-1010 W07-0715 H01-1035 H94-1028 J04-2003 D07-1005 P09-3009 N06-1015 H05-1024 W03-1001 W07-0711 J03-1005 H05-1023 C04-1060 P05-1074 Limited Context W99-0604 W06-3123 D09-1040 P09-1104 N07-1061 N06-1003 W07-0403 C98-2225 D09-1141 A88-1019 W02-1018 D09-1106 P02-1050 W07-0709 P05-1059 N09-1013 P02-1039 P00-1056 J03-1002 P08-1012 N04-1021 D08-1033 W09-0435 W08-0306 J00-1004 P01-1030 W07-0401 P07-1108 P06-1067 P03-1019 D09-1105 W07-0716 P05-1032 J95-4004 W08-0307 W05-0909 P05-2016 P01-1067 W06-1609 E06-1005 N07-1008 N04-1033 P03-1011 J97-3002 W09-1114 H05-1095 W95-0107 E06-1032 D07-1038 C04-1073 N06-1004 W05-0634 W04-3207 D07-1030 W06-3108 N04-1030 P96-1023 J94-4004 P07-1020 P02-1038 C98-1013 P03-1002 W04-3212 D07-1007 P02-1040 C04-1030 D08-1066 P06-1146 W04-3250 W04-2412 H05-1098 C04-1090 P96-1021 W06-3114 C08-1064 W03-1008 P05-1072 W04-3228 N03-1021 N04-1023 W09-0809 D07-1055 N03-1017 P05-1069 N06-1002 W04-1513 W07-0718 W09-0426 W05-0620 W00-0726 J02-3001 W09-2310 D09-1136 W02-2018 H05-1101 P09-1106 P02-1031 J05-1004 W02-1039 W09-0434 D08-1011 J96-1002 P06-1091 P07-1005 D08-1089 N04-4026 D07-1062 P05-1073 J99-4005 W08-0302 W08-0304 W03-1006 W05-0908 J04-4002 W06-1668 P07-1091 W06-2303 J92-4003 H05-1021 P08-1115 J07-2003 W08-0403 C04-1204 W05-1509 P04-1083 E09-1061 P05-1066 W06-3106 P03-1021 P07-2045 P07-1040 N06-2026 N01-1025 P07-1090 E09-1033 W06-3104 W08-0308 E03-1071 I08-2087 C08-1137 D09-1123 P06-1098 N07-1070 I08-1066 P07-1002 P01-1017 W03-1013 W05-1511 H05-1099 P08-1009 P06-1096 N06-1031 D07-1091 D09-1114 W05-1514 P06-1123 D08-1010 D08-1060 W07-2208 N03-1028 D09-1037 H94-1020 H05-1078 P02-1036 C92-2066 N03-1033 A00-1031 E09-1049 J99-2004 P09-1036 W08-0401 N07-1029 C08-1050 N04-1032 W96-0213 W00-1320 H05-1059 W07-0405 D07-1056 W03-0401 N06-1032 C08-1041 P06-1088 W97-0302 W05-1507 N03-1016 P08-1024 A00-2031 D08-1076 P02-1043 J97-4005 P06-2101 N09-1029 P09-1066 P06-1066 W02-1002 D07-1080 W06-1619 P00-1058 J01-2004 N04-1014 J03-4003 D08-1012 W07-0702 N09-1025 D07-1077 W07-0701 P02-1034 J07-4004 C04-1041 P99-1069 J05-1003 P07-1120 N04-1022 D07-1104 P05-1033 P03-1046 N06-1040 D09-1023 D09-1073 P05-1039 W98-1115 W07-0706 P00-1061 W09-0103 D07-1003 W06-3119 N06-1019 P09-1042 D09-1021 D08-1065 D09-1008 W09-0424 N04-1013 P97-1003 W02-1001 C08-1127 P92-1017 N06-3004 W07-2202 C08-1144 W09-2306 P02-1035 N06-1033 W04-3215 P05-1011 D08-1008 A00-2018 W06-1628 P04-1014 D09-1119 D08-1024 J98-2004 H05-1036 D07-1079 P06-1077 P02-1042 P01-1010 P09-1103 W05-1515 N06-1024 J07-3004 N03-1014 N06-1022 W06-1608 W08-0402 P05-1034 W04-2003 P95-1037 P99-1059 N09-1027 D09-1108 P06-1121 E09-1044 P07-1032 W07-1202 P08-1037 J93-2004 W00-1201 W06-1666 P09-1087 P08-1025 P09-1020 N09-1049 H05-1064 N04-1035 W06-1638 W03-1005 P03-1056 D08-1093 J99-4004 D07-1078 P03-2041 N07-1063 P04-1041 P07-1089 P01-1042 W97-0301 J98-4004 W06-1615 P03-1054 P96-1024 C08-1138 P08-1114 P09-1065 P96-1025 P05-1038 W04-3224 W06-3601 D09-1135 E09-1090 P02-1018 P04-1015 D09-1059 D09-1127 P04-1042 W06-2904 D07-1027 H91-1060 P08-1066 C02-1126 W09-1008 N09-1026 P08-1064 I05-1007 P05-1023 W07-2219 P05-1022 E03-1005 P09-2035 P03-1055 W01-0521 W06-1606 P04-1082 P06-1110 J04-4004 P05-1067 P09-1064 P03-1013 H05-1100 N07-1051 P07-1019 P06-2089 D09-1161 W05-1506 W06-3603 P07-1080 W05-1512 E09-1037 P05-1044 W05-1504 P04-1061 D09-1085 P06-1072 D08-1091 P08-1109 D08-1022 P07-2052 P08-1023 N06-1020 P09-1063 N06-1021 H92-1026 P08-1006 W04-3201 P04-1013 W05-1513 P06-1055 P08-1067 P05-1010 P07-1079 W08-2102 P09-1059 P06-1033 W09-1104 P06-2041 D09-1087 N07-3002 W05-1505 P06-1043 E03-1008 W05-1516 P99-1065 C08-1081 D08-1016 J08-3003 P09-1041 W07-2217 P09-1007 C08-1095 P05-1012 W07-2218 D07-1111 C96-1058 P08-1061 D07-1070 C04-1010 Revised Design: P08-1068 P09-1039 E06-1010 I08-2097 D08-1059 D07-1014 D09-1060 D07-1101 P07-1122 N07-1049 P05-1013 W06-1616 D09-1058 N06-2033 D07-1099 D07-1015 Information Criterion-based E06-1011 W06-2922 W07-2216 D07-1119 I08-1012 D08-1017 W06-2920 J08-4003 D07-1096 W04-2407 P08-1108 D07-1097 D07-1013 N07-1050 W06-2933 N09-2066 H05-1066 P09-1040 W06-2932
- 22. Connecting/Pruning Nodes Network Structure Learning Proposed Method Proposed Proposed Method Propo Learning in Collaborative Settings Network Structure Learning yt+1 yt+1 yt+1 yt yt yt input index input index input index input index input index input index input index input index Model 0.4 a)a) Adding training point causes many b) Adding training point causes many b) Ad output estimates toto change. output estimates change. out output feature index output feature index output feature index 0.35 SL outpu Traditional Collaborative Proposed Method Proposed Approach Proposed Approach Proposed Method 0.3 0.25 Proposed Approach Proposed Approach yt+1 2 Use ychanges in in the estimates ⇥yt yt+1 ⇥2⇥to P (yt+1 ) Learning in Black-box Settings 0.2 Use changes the estimates ⇥yt Figure 5: T1 =⇤yt yt+1 ⇤ and the yt+1 yt+1 y t+1 generalization error G d ) generalization error G (x(xd ) t+1 0.4 0.4 value that it tries to approximate ⇥G 0.15 yt y y 0.35 (Section Section 3.1). Most importantly, yt t t 0.1 0.35 0.3 high values of ⇤yt yt+1 ⇤2 should cor- input index input index where yt+1 are the estimates after d , y, ) ) where yt+1 are the estimates after (x(xd d ydw input index input index 16% 0.3 84% respond to high values of ⇥G, since those are the points that are likely to be 0.25 0.25 P (yt+1 ) chosen. 0.05 P (yt+1 ) 0.2 0.2 Figure 5: T =⇤y y ⇤ and the 1 t t+1 0.15 Figure 5: that it tries to t+1 ⇤ and the ⇥G value T1 =⇤yt y approximate 0 0.15 value(Section tries to 3.1). Most importantly, that it Section approximate ⇥G y y high values of ⇤yt yt+1 ⇤2 should (Section Section 3.1). Most importantly, cor- 9 0.1 16% 84% high values of to high t+1 ⇤2 should cor-since respond ⇤yt y values of ⇥G, Proposed 0.1 yt+1 A−optimal 16% 84% respond toare thevalues of ⇥G, likely to be those high points that are since D−optimal thosechosen. 8points that are likely to be are the 0.05 E−optimal Transductive chosen. 0.05 Random Optimal Figure 3: Distribution of the ⇥ estimates 0 y y ⇥ 0 9 7 Proposed yt+1 in relation to the estimate yt and the y y y t+1 9 A−optimal D−optimal Proposed E−optimal Mean Squared Error y t+1 8 A−optimal Transductive f (x) f (x) true value y.xFigure 3: Distribution of y estimates D−optimal Random 6 E−optimal x 8 Optimal yx the Transductive x Random yx = yx = 7 ·x ⇥ x Figuret+1 in relation to·the estimate yt and the y 3: Distribution of the estimates 7 Optimal Mean Squared Error true value y. ˆ yt+1 in relation to the estimate yt and the 6 5 a)a) Adding training point causes many b) Adding training point causes few Mean Squared Error yt+1 true value y. (a) White (x) Model. f Box (b) Black Box Model. Adding training point causes many b) Adding training point causes few 6 x 5 yx ˆ yt ˆ y y* yx = ·x t+1 4 5 ˆ yt+1 output estimates to change. output estimates toto change. output estimates change. ˆ yt y* c Figure 4: For Black Box models only the inputs and outputs a b 4 output estimates to change. a b * c ˆ y y t 4 (b) Black Box Model. c 3 a b 3 are accessible, internal workings areFigure after4:after after training unlike for Figure 4: not accessible, point yFigure the training point y the training is added to the training set point 4: y the (making the 3 is added toadded to the training set (making the the is the training set (making 2 1 2 3 4 5 6 7 8 9 10 models only themodels [11].outputs of trainingofpoints points equal to t+1). White Box inputs and number Training Set Size 2 2 1 2 6 3 4 5 6 7 8 9 10 1 2 3 4 5 7 8 9 10 number number of training pointsequal to t+1). equal to t+1). Proposed Approach Training Set Size Training Set Size kings are not accessible, unlike for Figure 6: Evaluation of active learning Figure 6: Evaluation of active Error: lower criterions (Mean Square learning criterions (Mean Square Evaluation at active learning values are better, values are different of Proposed Approach yt+1 yt+1 Figure 6: Error: lower yt+1 2 estimate improvement in the the statistical significance level of 95%). values are better, values are different at criterions (Mean Square Error: lower Use changes in in the estimates ⇥yt yt+1 ⇥2⇥toto estimate improvement in the Use changes the estimates ⇥yt yt yt input index input index the statistical significance level of 95%). values are better, values are different at a) Adding a training point yt+1 yt b) Adding a training point the statistical significance level of 95%). generalization error G d ) generalization error G (x(xd ) influences many output es- input index influences only a few output
- 23. t+1 f (x) 1 + x A x ⇤G p J 1 ⇤ 1 ⌅ A = X⇤ X + I A = X⇤ X + I xx A A X⇤ X = ⇥i i ⇤ , Conceptual Justifications (A + x x⇤ ) 1 = A f1 i (x) . ⌅p ⌅ p 1 + x⇤ AJ(⇥) = ⌅⌅t 1x y ⌅1 yt+1 ⌅ .⇤2 i=1 = ⇥ ⇤ + I. = ⇥i i ⇤ + I. i 2 ⇤ t+1 = (A + x⇤x⇤1 1 (X⇤ y + x 1y )⇤ 1 ⇥i ⇤ ⇥1 ⇥2 . . . pi ⇥im i ⇥m+1 =i=1 . = ⇥d = .. B T1 V ) ⇤ A xx 0 A X y i=1⌃ p m a ⇥ 1 1 ⌃x⌅ ⌅ . 2 ⇥2 ⇥ X⇤ X (A + x x ) ⇤ X ⇤G A X y ⇤ 1 y= i a⇥ 2 (A + x x ) 1 X⇤ y + (A + x x ) x1y . x A E G = = + B + V, + ⇤ 1x A ⇥ xi ⇥ i . JO 2 BJ V ⇤ 1 x⇥ A x⇥ = a ⇧ ⇥ x⇤ A 1 x⇥ = a i=m+1 JO i=m+1 JO ⇥ A 1 x x⇤argmax⇤J(⇥). A 1X y 1 p⇧p ⌅ ⇥2 ⇥ (A + x x⇤2) 1 X⇤ y = A 1 X⇤2y n i=m+1 x⇥ x⌅ i 1 2 x⇥ x⇥ 1 i=m+1 ⇥ i ⌅ ⇤X + I E G = +J(⇥) = ⌅⌅ B + V, y {⇥.i }i=1 1 + x A 1 x ⇤ A = XX⌅ X XX { { p 2}p }i=m+1 a2 x t ⌅ yt+1 ⌅ i a i i=m+1⌅ J = ⇥ ⇤ (A + x x⇤ ) 21 x y = A 1 x y ⌅ A 1 x x⇤ A %-1#x y '# ,# ⌅ p JO = 2 . X X⌅ X XO (a + 1) . i} 2{ { {⇥i }n ⇤ 1 1 ⇤3 1 / +$, ⇤#)$* 1 . . .,/')-'##,# JO JO = ⇥i i ⇤ + x⇥ (a + 1) i I. x A x x 1A + x'%' A x51 ⇥ B= E f (x) + x i=1 ) = A 1 (A g(x)x q(x)dx, . ( JO i=1 JO ⇤ ⇥ ⌅2 1 1A x y1(1- ⇤ + x⇤ A 3 1 x 2 D argmax J(⇥). A x x !#$%' a a JO a x⇤ ) 1 ⇤y B = + x E f (x) xg(x) = A x y (A q(x)dx, 1 ⌅2 ⇤ 1x 3 . 3 0$1 0 x⇤ A 1 x⇥ = a a JO a JT JT D V⇥ = E f (x) E f (x) q(x)dx. x A 1+ 14 16 ⇥ JO JT JT ⇤ ⇥ D ⌅2 ⇥ ⇤ ⇤ q(x)dx. ⇤ E f (x) =A1 1 x y A 1 x x⇤ A 11 x y ⇤ 1 1A ⇤ x x A ⇤ 1 ⇤ X y (1 + x⌅ A⌅ A ⇥1 x⇥ )2 1 1 ⌅ (1 +⌅ xa2 ⇥ x 1 )2 p f (x)t+1 + x tx ) X⇤ y = A 1 X⇤ y A x x A X !#$%'(!)*+, y ⇥2 1 p⇥ V =E (A JO = x⇤ i 2 2 ⌃⌃ ⇥ . a= 1 x⇤ ⇥ 2 ⇤ 1x ⇤ ⇥ ⇤ ⇤ 1 a= Figure 7: Decomposition of generalization error G into model error C, bias A variance V , where g denotes 1 + x A 1 +2 A ⇥ D ⇥ x 1 x 1 + x B, and 2x (a T J1) ⇥i + x⌅ A⌅ A t1 x . ⇥i.+ ⇥ ⇥ 1 i ⇥ ⌅ ⌅ J+ T i=1x x t 2 1 ⇥ yt ⌅ f⇥ ⇤ optimal function, ⇤is a⌅⌅1 1function⇤i⇤ are the1 ⇤ functions from a slightly di erent training set. 1 J(⇥) f= learned yt+1 ⌅’s = 1 X learned t t+1 , 1 ⇤ i=1 ⇥ ⇥ ⇤ =A 1 x y A x x A 1 + x AA x x xx A X y A x x AJO X y 1 xy A x 1 ⇤ ⌅ m xt ⇤X x ⌅ m ⇥2 xt ⇤X x 1 ⇤p ⇥2 ⇤ 1 ⇥2 1 ⌅ 1= ⌅ p t =A x6.2.1 Parameter Change-based 1 ⌅ 1 y = x⌅ A⌅x⇥xt1i ⇤ + a x⇥ i x ⇥⇥i +i . + 1 +AL 1 + x⇤ At1 ⌅x items thata 1likely toainfluence thex the most. 1 1 + x⇤ A 1 x ⇤ x(Settles et⇤al., 2008b) favors 1 A x t+1 + x⇤ A 1 model ⇥ x A xt ⇥i + i=1 ⇥ i=1 i=m+1 i=m+ 1 ⇤ changes A model’s parameters⇤ yfor the better, i.e. approach the optimal parameters, it x x 1A x Parameter Change-based are ⇤ ⇤ 1 1 + ) x xis then beneficial to select the + that A x are . 1 X⇤ y (A + x x x A y11x Assuming x A⇤ 1 xin an itemA ⇤has the1greatest impact on the model’s parameters: A = A x x ⇤ t )21 x x x A (y thaty 1 ⇤p =A 1 x y p ⇥ ⇤ ⌃⌃ p ⇥ ⇥⌅ ⇥ ⇥ 1 1 y ⌅ = = x⇤ ⌅ J(⇥) = ⌅⌅t yt+1 ⌅1 + 1 X A ⇤1 x ⌅x 2 1 t+1 . , G = (x )⇥ X A I,x , ⇤ = A 1 X⇤ y. ⌅ A⌅⌅t1=t = ⇥ x⌅ x⌅ xt x⌅ A 1 + = ⇤ + y Y1 T Tt⇥(x ,y)), ⇥ change X x x⇥ x A x 1 px (22) ⇥ 1i i i=m+1 a + x tA a E L( a a= ⇥ x⇤i=1 i=1 ⇥ i 2 ⇥ i t ⇥i +i ⇥i + ⇤ I where T ⇤ t+1 ⌃m⇥ + ⇥ A 1 q(x) x⇤ ⌅ t )⌅ parameterare the model’s parameters estimatedyfromanthe current training ⇥TtrainingTset T,y)and Li=1model’s m ⌃⌅i ⌅⇥ x⌅ ⌅⇥ p⇥ 1 1 + 1 1 p ⌃x⌅ ⌅⇥ x⌅ ⌅⇥ . p(x) ⇤=x (y estimates after a hypothetical rating of set , and ⇥(x item xa 0 added to the 1 is , are the a m the loss is = ⌃p ⇥ = ⇤ 1x t . t+1 1 function that measures x y erences between the parameters. ⇤ y A xx A ⇤ 1 the di 1 A xx A X ⇤ 1 ⌅= ⌅ = ⇥2x⇥ x⇥ i t xti ⌅ + ⇤ + 1i 1 ⇥i ⇥ + ⇥ 2 i ⇥ xi ⇥ i t xt i i (x) ⇤= q(x) 1 + x =A 1 x y A ⇤ ⇤ A ⇤ 1 ⇤⇤ t x⇥ i=1 i=1 + i x⇥ i i=m+1. i=m+1 x⇤ A= x A x 1(y ⇤ x 1 t ) . i t+1 t 1 ⇥i + ⇥ ⇤ 1 +Variance-based decomposed into + x A xmodel error C (the di erence between the opti- yt+1 approachtthe error is ⇥ 6.2.2 ⇤ y X ⇤ t+1three (A 1 xx x ) (X y + x y ) ⇤ components: ⇤1 1⇤ ⇤ i=1 i=m+1 1 ⇧p⇧ 1 i=m+1 x⇥ x p ⌅ = X X +x⇤function approximation Agiven1 + xAA x +the A function f ), bias B (the di erence between ⇧ n ⇤ A 1 ⌅ 1 + malI, 1 x In this A = 1 x ⌃ x⇤ g, 1 x current model, and ⇤ true X y the 1 ⇤p ⇤ ⇥2 i=m+1 ⌥ =A i )x y 1 ⇥2⇤ the current approximation ⇥ ⇤ an i=m+1 x⇥ i ⇧ n min q(x ⇤ A 1 xIn(y A 1 x fhave:) optimal one g), and+ x AV⇤(how1much y + (A approximation f 1 x y . xwe andt = (A variance 1 x X⇤ the function + x x⇤ ) ⇤ x ) ⇥ I ⇤ ⌅ ⇥ ⌃ ) + x words, f (xivaries ).yi ,⇥ 1 0 other 1+x a ⇤ ⌥ q(xiyt+1 p(xi ) = X 2 ) ⇤ yt ⇥ . G = C + B + V. (23) p 1 ⌃⌃ ⌅ ⇥ ⇥⌅ ⇥ ⇥ p min A i=1 f (xi1)x (yi A y x t One1⇤ ) x⇤ A 1 x + ⌅t , ⇤ solution (Cohn et al., 1996) is to minimize the variance component V of the error by assuming that the x⌅ A⌅ A t1⇥t ⇥ 1 ⇥ x ⇥ 1 x x x ⇥ x ⌅ xt x ⌅ , , ⇥ i i t i i p(xi ) = . bias component becomes negligible (if this assumption is not satisfied then this method may not be e ective). i=m+1 J( ) = variance ⇤a number of⇥2 i=1 i=m+1 ⇥ ⇤ ⇥ 1 1 + x⇤ A Theret+1 the model’smethods proposed that aimJOselect training inputs seeks to selectcertain measure points 1 x are y ⇤ ⇥⇤ of yt parameters. The A-optimal design (Chan, 1981) for reducing a training input of the to JT JT JT ⇤ 1 as ⇥2 1J( ) = (y x⌅ ⇧ )2 x⌅ A 1 X⇥⌅ X⇥ A sox to minimize the average variance of the parameter estimates, the D-optimal design (John x⇤ A 1 A 1 x Draper, y 0 x⇤ ⇤ 1975) seeks to maximize the di erential Shannon information content of the parameter estimates, and the t ⌅ 2 ⇤= al., 2006)+ x tox⇤ ⇥ 1 = 1 1 (1 + x A 1 x )Transductive Experimentaltdesign (Yu et ⇤ (A 1seeks⇥⇤ )representative training points that may allow . find A ⇤ ⌃⌃ ⇥2 J( ) S= ⇥⇤t+1 yt ⇥2 0= retainingcomponent,1also 1= 1 ofaccount Aexistense of(x⇤model1inx(Sugiyama, 2006),+ x A to the JT JT 1 J y ⇤ = variance mostA the x (y into xxtesttpoints. X ALthe AAerror2component.1 in addition x ⇤ of information ⇥ ⇤ ⇤ takes the ⇤ ) ⇤ The Xmethod ) x. x⌅ A⌅ A t1 xt ⇥2 ⇥ x 1 x = JR , yt+1 y⇥= X x A x ⇤ ⇤t 1 + the .O = J ⇥ ⇥ , xt ⇤X x ⇥ J( | y = r) JP 2 = 1 1 + x⇤ A 1 x ⇤ Image Restoration-based (1 + x⇤ A 1 x⇥ )2 xt ⇤X x ⇤⇤ ⇤ ⇤ d d ⌅2 ⌅ y x t 6.2.3 ⇤ ⇥ ⇤ ⇥2 1 ⇥⇤problem of predicting the user’s preferences as one of image restoration1(Nakamura ⇥ 1 ⌃⌃ 1 ⌃⌃ ⇥ ⇥⌅ ⇥ ⇥ 2 = ⇤) , 1 ⌅⇧ 2 x A X XA x. It is also possible to ⇤ treat the (x⇥ A x⇥ ) 1 2 xt ⇥X x ⇤ 1 x⇥ A ⇤ xt ⇤ A xx A 1 ⇤ X y ⇥ ⇥ 1 x⌅ x⌅ xt x⌅ . . JR = (y1 + x t A x x JO = , J⇤ of 1 ⇤ Abe, 1998), that is, based on our limited knowledge = a user’s preferences (a1 (A + x x ) X y = A X y . partial picture), we try to restore ⇤ user’s likes and dislikes.TThe AL task is+ x⇤ A 1 x )training points that would ⇥ i ⇥ i i t i the complete (1 + x the picture of A ⇥ 1 x )2 ⇥ (1 then ⇥ select the 2 to ⇥ 1 + x A 1x ⇤ xt ⇤X x xt ⇤X x i=m+1 i=m+1 J( ) P (y = r)J( JS = x⌅ A 1 X⇥⌅ X⇥ A 1 x , best allow us ⇤ ⇥2 to restore the “image” of the user’s preferences. It is interesting to note that this approach satisfies ⇤ 1 r 2 the desired properties of the ALx A outlined in Section 2. For example, if a point already exists in aJ methods xt region, J( ) = ⇥⇤ y ⇤ y⇥ xt ⇥X x ⇥ JO JT JT T x⇥ x⇥ X⌅X⌅ X JP = (1 + x⌅ A 1 x )2 .t+1 tT thenJ = sampling neighboring points the image in that region could likely be restored. This approach also without . ⇤ ⇥ ⇥ ⇤ edges of 2 may favor sampling close to the (1 + x A X X 1 x ) image components (decision boundaries). X x⇥ ⇥
- 24. (Section III-A1). Size of the node reflects degree of the node (scaled logarithmically). Connecting People relation to number of alumni nodes differs. Stanford University has a significantly higher ratio of companies per alumni in leadership roles, followed by Harvard, trailed by Berkeley, and subdued edges in Figure 2. The distance from the universities University Alumni to the cloud of ‘financial’ clusters also varies. In particular, Stanford MIT Stanford and Berkeley are rather close to the financialPerson Other Nodes Financial Org. Company cloud. t t fi Collaboration Networks Harvard MIT. A high ratio of company nodes indicates that alumni This may be explained by the geographical proximity of these Berkeley University b have been involved with multiple companies - either through universities to one of the largest sources of venture funding – Size (Centrality log-scaled) l employment, advisory or investment activities. In addition, Silicon Valley. While universities themselves are not embedded o the number of highly connected alumni (large nodes with within the financial clusters, a noticeable proportion of alumni f many connections located on the perimeter) differs significantly are deeply connected within the financial clusters by having N University Alumnus between universities (we further explore this in Section III-B). Company direct or indirect relations with multiple financial organizations. a One particular characteristic of highlySize (degree log-scaled) connected alumni stands Stanford has the largest number of alumni connected to the a out, namely their collaboration patterns. Stanford’s densely financial cluster, followed by Harvard (even though university T connected alumni are highly likely to collaborate with fellow itself is relatively distant from the financial cluster); followed a alumni (indicated by the company nodes being pulled away by Berkeley, and only a few alumni from MIT. The proximity t from highly connected alumnus towards other less-connected between alumni and their alma matters appear to differ signif- l alumni in the center). In the networks of other universities, the icantly. Berkeley alumni tend to be clustered together, MIT to t collaboration of highly connected individuals with their fellow somewhat lesser degree, and Stanford and Harvard alumni are N alumni is evidenced but to a lesser degree. rather dispersed. Proximity between universities differs as well. Harvard alumni appear active in leadership positions in Stanford and Berkeley are close together (many alumni hold n technology-based startups, even more so than MIT alumni leadership positions in the same companies). One of the likely v (Figure 1b vs. 1c). A possible explanation for the relatively explanations for this network proximity is the geographical t lower level of MIT alumni may be attributable to the focus of proximity of both universities to Silicon Valley where many of b (a) Stanford University. (b) Harvard University. (c) MIT. (d) UC Berkeley. this dataset on leadership positions in the organizations. While the investment firms and startup companies are located. Harvard S Figure 2: Inter-University NetworkasNetwork and MIT do not appear to have (between relations with other Inter-University strong Stanford, Harvard, p Figure 1: Intra-University Networks. Networks of the above universities are expanded in a breadth first manner so in the depth engineers play a key role, they often do up to a technology universities in these settings. MIT, Berkeley) (Section III-A2). Network is obtained by start- U of 2, (showing university, alumni and companies they are associateddevelopment capacity rather than in the leadership activities) that with through employment, investment or other positions ing with the nodes of the above mentioned universities and t (Section III-A1). Size of the node reflects degree of the node (scaledare visible in public relations communications. Some support logarithmically). for this explanation may be seen in the Figure 2 relatively large performing a breadth-first expansion up to the depth of 3. p distance between Microsoft and University of Washington, even e though a large number of engineers atUniversity Alumni Other Nodesindeed Microsoft are w relation to number of alumni nodes differs. Stanford University has a significantly higher ratio of companies per alumni in from University of Washington. Stanford Financial Org. level entities/nodes that are linked to previous levels including MIT 2) Inter-University Network: A graphic representation of Company financial organizations, company employees, etc. Since we B leadership roles, followed by Harvard, trailed by Berkeley, and Harvard Person MIT. A high ratio of company nodes indicates that alumni the alumni-based inter-relations between Berkeley universities, shown University are primarily interested in relationships among alumni, all of have been involved with multiple companies - either through in Figure 2, was produced as follows. Four universities were Size (Centrality log-scaled) other entities are faded out, except for the above mentioned n employment, advisory or investment activities. In addition, selected for analysis: Stanford University, University of Califor- universities and their alumni. In addition, we glimpse at the o nia (Berkeley), Harvard University and Massachusetts Institute relations between alumni and investment firms (a very impor- n Betweenness Centrality Betweenness Centrality the number of highly connected alumni (large nodes with Betweenness Centrality Betweenness Centrality many connections located on the perimeter) differs significantly of Technology (MIT). From these nodes we have performed tant factor for entrepreneurism). Therefore, nodes of financial l between universities (we further explore this in Section III-B). breadth-first expansion up to the depth of three: 1st level organizations are not faded out. q One particular characteristic of highly connected alumni stands being alumni of these corresponding universities, 2nd level Two distinct groups – universities (in the lower left corner) n out, namely their collaboration patterns. Stanford’s densely are companies with which alumni have relations, and 3rd and financials (in the upper right corner) – are visible in the S connected alumni are highly likely to collaborate with fellow 1 alumni (indicated by the company nodes being pulled away D from highly connected alumnus towards other less-connected p alumni in the center). In the networks of other universities, the i collaboration of highly connected individuals with their fellow alumni is evidenced but to a lesser degree. e Harvard alumni appear active in Alumni Indexleadership positions in Alumni Index Alumni IndexIndex Alumni University F (a) (a) Betweenness Centrality Alumni Nodes in so Networks (University and (b) Betweenness Centrality Alumni Nodes in Networks (University andand Betweenness Centrality of of even more Networks (University and (b) Betweenness Centrality of of Alumni Nodes in Networks (University technology-based disconnected). startups, Alumni Nodes in than MIT alumni Other Entities (Company, People, Financial Org.) m alumni nodes are are alumni nodes disconnected). alumni nodes are are connected). alumni nodes connected). (Figure 1b vs. 1c). A possible explanation for the relatively Size (Centrality degree log-scaled) t Figurelevel of MIT alumni may be attributablealumni nodes thethe network (connected thethe university node (b)), and (disconnected Figure Betweenness Centrality (y-axis) of alumnithe focus of network (connected to to university node (b)), and (disconnected lower 4: 4: Betweenness Centrality (y-axis) of to nodes in in t University-Company Network Figure 3: Universities within the Business Network (partial c this dataset on leadership positions The x-axis corresponds to alumni ordered in descending order of centrality for each of the from thethe university node (a)). inx-axis corresponds While from university node (a)). The the organizations. to alumni ordered in descending order of centrality for each of the
- 25. — — —- Generalized Assignment Problem —- Connecting People Contents —- Generalized Assignment maximize Generalized Assignment Problem p, choose a Problem — Objective: given a paper M given a Assignment Problem collectively possesses Generalized a fixed size s) that Objective: ⇥ M (ofpaper p,given a paper subject to a gro —- the most expertise about p: a groupa ofp, choose Objective: choose Objective: given a paper p, choose group of experts group of experts experts R(M Expertise Finding Group Formation M ⇥ M (ofgiven ⇤papers) that fixed sizeof experts M ⇤ M a fixed size (ofchoose a group possesses collecti M a M p, a collectively s) that Objective:⇥ M (of a fixed size s) that collectively possesses M size s) most collectively possesses the most expertise about p: the that expertise about p: the most expertiseexpertise about P the most about p:R(M, p) = p: maximize m2M r (m, p) Task Description Task Description 6'1'.% |M = R(M, p) ⌅= P subject to maximize R(M, p) = m2M r |(m,sp) P maximize R(M, p) = m2M r (m, p) P P !#$%'()*+*,-).(./) maximize R(M, p) |M | = s r R(M, p) = subject to maximize (m, p) = m2M 7%2*.*.1% Limitation X |M | = s subject to Group 896').*.1% subject to subject = s r (m, p) . |M | to Expertise⌅=is not additive: |M | R(M, p) X m2M — Limitation – R(M, p) ⌅= X p)r.(m, p) . R(M, p) ⌅= r (m, Training Set:— = (XT additive: iExpertiseT Estimation m2M 4'#')(5')% Expertise is not , YT ) = {(x , f (xi ))xi 2X } Limitation T Group m2M Training (and = (XT , Y Y ): RT Function learned from XSet: Tcorresponding ) = {(xi , f (xi )) Limitation – — b !#$%'#()*+,-.% Expertise is not additive: TAssumptions T T bT ) = L(RT , f ) b Expertise is not additive: Generalization error: G(R ⇤ Group FunctionEstimation from XT (and correspon Expertise learned /'010%('2*(%++')3% – – Assumptions M—- athors of paper p; so p are Group Expertise Estimation Generalized Assignment Prob Generalization error:T0, where =L(⌥ andf ) b b ⇤ R(M, p) R G(R ) =M1,RT , p = Researcher Profile Researcher Profile Researcher Profile Mp are Expertise EstimationG(= ) ⇤ T Group — — min Assumptions athors of paper Xp; ⇤ so R(Mp , p) ⇤ b M (-.7)*:;7'%$.-='1'% R(M, p)of—- where M ⌥ R(M ⇤ ,Ø. given a paper p, c —- Assumptions — Objective: = 1, and T Knows — p are athors M ⇤ = 0, paper p; so Mp = p) Knows Knows g: optimalare athors Assignmentp; so R(Mpa p) ⇤min G( b t Generalized(in the ⇤ ⇤ function sollution space) ⇤ , Problem of paper ⇥ p R(M, p) = 0, where M ⌥ MpM Ø. ⇤M (ofM ⌥Problem RT ) —- Mp—- GeneralizedR(M,=choose a fixed Tsize s) = Mp 1, R(M, = and b learned function given aM ⌥ Assignment group. of experts R: R(M, p) = 0, where paperProblem ⇤ Objective: Assignment pp, p) = Generalized M Ø. X Generalized Assignment Problem most ⌥ Mp Mp about p: the M expertise⇤ ⇤ b M ⇥ M (of fixed size p, choose a⌥groupchoose a Objective: given a slightly a paper Mp possesses Objective: given di⇣erent training Ri ’s: learned functionsaR(M,paper s) that collectively of experts p, Objective: given a paperfrom a p) = a group theexperts set.spac M p, function p) = of . sollution choose optimal (of p: (in the g: M use Learn R group expertise. B (ofV most expertise that collectivelycollectivelyestimate + ⇥ and ⌘⇥ M abouta s)and⇤ size ⇤ to . possesses EG = Learn abfixed (of aitfixed sizeb fixed use its) that colle M R C to R(M, Mp + M size s) estimate that possesses M ⇥ M ⇣ Mp b expertise the = R R(x)useg(x) mostfunction aboutP B most expertise learned expertise 2 Learn b and R: the estimate group expertise. maximize E bthe mostabout p: about p: — — p: R b it to it to YT R ’s:2 T = (XT , functions (x= P T ⇣ Training iSet: use maximize) =R(M, fp)expertise.}r (m, p) b ⌘ Learn R and learnedestimate{(xi , fromxia slightly di⌘ group i )) 2X maximize Set: p) = (X Y, Yto) = { Training P T subject T b — b b m2M V = R ER(x) Function = = B from X p) =R(M,i 2XT= | = T r (m, p) — Training Set: EG (XT , YsubjectTto, f (xicorresponding s T ): RT T learned +)V {(xC =+ maximize T R(M, i(and ))x r|M m2M (m, p) } (x)) ⇣ maximize R(M, p) = ⌘ m2M Generalization X subject TT))= {(xR f (x )) R s X (and Training Set: = Function learned from C = (g(x) fsubjectTto T (XT , Ycorresponding iYTxi|2X T} G( b |M |2i , Function learned from error: Challenge= T , f )|M): = T T b (andRto = L( b s 2 b b Function = ER(x) b b y corresponding G(): R) = Challenge Use changes B the estimates Xyg(x)t+1 to to from T (and in learned RT ) = L(RTsubject error: YTimprovem b Generalization estimate RT T |M Generalization error: G( additive: , t bf ) 2 b Challenge Generalization not Expertise ⇣ Challengeis ) generalization error G(x error: Expertise RT not additive: G(RT )⌘2 L( is ,bT)) b = min b RG( f b Expertise estimates b G( V = b where y is are the isR additive: R ,T) was not Expertiset+1 not additive: ER(x) (xbTy )X added to the train X after min !#!$% R(M, p) ⇤=G(RT ) r (m, p) X m g: optimal function (in XT number ofbtraining points is eq . the 2 min space) Gt : generalization error when the sollutionX C =function p) f (x)),space) added. tor p) R(M, ps X m2M b (g(x) sollution yp)Tr (m, m2M the ⌅= . R(M, X ⇤= R(M, (m, training g: t+1 : R: –learned (in the after (x m2M is generalization error ⇤= G optimal function R(M, ) p) p) !#$%'#()*+,-.% m2 ⌅G = Ritoptimal :functiong:– in insollution space) (in the t+1 R: (in the a b learned–functionchanges optimal functionerrortraining set b Group Expertise Estimation estimates y g: ’s: learnedimprovement the slightly different G Use Gt+1 functions from generalization bt y sollu b – !#!$% — EGGroup Limitation b learned V + =B function Estimation Ri ’s: Expertise+ExpertiseR: slightlyExpertise Estimation R:Assumptions C b b generalization⌘error G(x ) b Grouplearned learned functions learned function ⇣ functions from a different training set. ’s: b ExpertiseGroupa slightly different training se RiM ⇤ are Estimation 2 is not additive:, p) = 1, and from B= ECAssumptionsg(x)bpaper p; so R(Mp EG = B + Vp +R(x)⌘ y athors of ⇤ ⇣ EG = B + V 2 C min G the ⇤estimates after (x y aret+1 M max ⌅G. Assumptions ⇤ where+bt+1 i ’s: learned functions from ,a s RAssumptions ⇤ , p) = 1, and M⇣ are athors⌘2 x paper =p = R(Mp b R(M,g(x) 0, where MR(M ⇤ , xp)Ø. p) = of ⌃ p; so B ⇤ are R(x)p bof – generalization perror = 1, and⇤ numbe Mp= EathorsGb : paper p;⌘2 ⇤ so V = RR(x)b 0, where Mare + = Ø.when M E ⌘ R(x) EG p⇣ B ⇤ V + M ⌃ the B R(M, tp) = g(x) ⇤ M= ⌃ MpathorsC of ⌘ p E paper p; R(M,⇣p) = 0, where M ⌃⌘ Expertise Estimation 2 . b = (g(x)Group 22 p R(M, b = 0, where (x , yM ⇤ 2 M = Ø. R(M, p) = ⇣Gt+1 : generalization error M ⌃⇤M ⇤ ⌃ ) is C b V = R ER(x) f (x)) B = ER(x) ⇤ g(x) p p) after M Mp V = R Assumptions R(M, p) p b ER(x) b M ⌃ M= p C = (g(x) changes in the Use f (x)) ⇤ Gt estimates improvement in. genera ⌅G2 = ⌅G2=JGt+1 : ↵bt ⇤ yt+1 ↵2 Mp estimate imp R(M, + Kp)⇣ = y b . ⌘ to⇤ C Learn R and use it to estimate2paperexpertise. R(Mp = (g(x)Mp (x)) athorsb ofMgroup 2 so R(M b f are ⇤ — b semantic and structural to p; Use generalization error G(x )= t Rt+1 ER(x) 2 ⇤ improveme p changes in the estimates yy yyt+1 ↵ b estimate = ↵b 2 J V bt b b Use b R(M, estimates y both features. and yt+1 in Kp) =group ↵bt ⇤M y ↵ to = Ø. generalization error G(x ) the =2 y0, where(x t+1t+1was added to th bwhereuse R are the estimates after group⌥) Mpestimate imp Learn and use it to estimate y , y expertise. Learn RUse changesto estimate b it expertise. b generalization errorstructural(g(x) t , y f (x))2 ⌦ training point y b Use ensemble predictive structural features. model. Use both semantic ) when the number of G(x and features. b b semantic estimates = (x , y ) was added to the train t+1 Use both t : generalization error G where yt+1 are theand C after where yt+1 are predictive model. , y , yis) was added max — b Use ensemble Learn b in ) xR(M, = is tra Use ensemble predictivethe Use changes (x training t+1estimat estimates after min G to M th Gt : generalization error model.the after (x of useadded to the x ↵ —- Gt+1 : generalization error Y )R and, fthe it to p) = equ when number points estimates GtTraining Set: T = (XTwhen =in numberi ))xi 2XTerror :— Gt Gt+1 : improvement {(xi ⇤(x of training point ⇤G generalization true output the semanticnot training se = sinceerror error (xboth is generalizationaccessible. , T } Training Set: T fromTX T (and{(xy, f (x ))xand }struct Use , y valuesiy) isto 2X —t+1 :: generalization the after ) #$'()(*+,'%),!#!$% G⌅G estimate generalization(x addedare the Y G Function learned= error, Yf (x= error } added to ): RT (X ) G(x i ) Training Set:Ggeneralization= {(xi afteri ))xcorresponding T T the tra : ⇤G = Gtt+1 T =: (XT ,⇤YT ) Use ,Tgeneralization ierror , improvement inensemble Tpredictive mode b t+1 i 2X b
- 26. Connecting Models Operationalization of complex conceptual models Motivation Internalization hFp://cmap.ccdmd.qc.ca/rid=1225215801935_1377957481_826/
- 27. Model Diversity Table 10.1. Motivation Terms: Author by Factor Matrix Author Wlodkowski Paulsena Donald Keller MacKinnon Panitz Feldmanb Nuhfer Farmer Theallc Pintrich Forsythd Chickeringe Factor Inclusion X X X X X Community X X X X Climate X X Ownership X X X X X X Attitude X X X X Affect X Interest X X Awareness X X Attention X Enthusiasm X Meaning X X Relevance X X X X X Value X X X X Competence X X X Empowerment X X X Confidence X X Expectancy X X X Leadership X X High expectations X X X Structure X X X X X Feedback X X X Support X X X Satisfaction X X Rewards X X a Paulsen and Feldman (Chapter Two). (Theall Franklin, 1999) b Feldman and Paulsen (Chapter Seven). c Theall, Birdsall, and Franklin, 1997. d Forsyth and McMillan, 1991.
- 28. Complex Modeling Typical Approach Feature Extractors Features Sub-Models Sub-Models Aggregation Selection Conceptual Conceptual motivation relatedness Expct Val T Expct Val T Text (tweets) label A1 .……………… sentiment Cogn Diss T Cogn Diss T ... ... .……………… A2 .……………… Textual Classifier Computational .……………… part of spch M1 M1 A3 .……………… dependency M2 A4 M2 ... ... Dataset Construction Feature Extraction Sub-Modeling Ensemble Construction
- 29. Summary --------- --------- --------- --------- --------- AI --------- --------- --------- --------- --------- --------- --------- --------- --------- --------- Networks are crucial for learning • use AI methods to construct / opTmize networks – connecTng contents – connecTng people Concept Extraction concept concept concept concept concept concept concepts concept concept Group Formation documents – connecTng people contents tasks – connecTng models context (documents) Semantic Mapping concept concept concept concept Influence Estimation concept concept concept concept concept concept concept concept concept concept interaction log concepts concept concept Knowledge Level concept concept Estimation concept concept concept concept concept concept concept concept concept concept concepts