SlideShare a Scribd company logo
Introduction   Thesis   Problem        Proposed Approaches   Prior Work   Timeline




         Active Communication for Distributed Learning

                             Avishek Saha
               School Of Computing, University of Utah

                                  PhD Proposal
                                  4th May, 2011




                                                                              1 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                               2 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                               3 / 29
Introduction       Thesis     Problem    Proposed Approaches   Prior Work   Timeline




Large scale learning




         1     Machine learning on large scale data

         2     Data size surpasses the limits of a single core

         3     Distributed learning on multiple cores




                                                                                4 / 29
Introduction        Thesis    Problem     Proposed Approaches   Prior Work    Timeline




Challenges




         1     Parallelization of serial algorithms/new parallel algorithms

         2     Scheduling issues in multiple cores

         3     Combining results from multiple cores

         4     Communication bottlenecks in distributed learning




                                                                                  5 / 29
Introduction        Thesis    Problem     Proposed Approaches   Prior Work    Timeline




Challenges




         1     Parallelization of serial algorithms/new parallel algorithms

         2     Scheduling issues in multiple cores

         3     Combining results from multiple cores
               ยง                                                         ยค
         4     Communication bottlenecks in distributed learning
               ยฆ                                                         ยฅ




                                                                                  6 / 29
Introduction       Thesis   Problem    Proposed Approaches   Prior Work   Timeline




Observations


         1     We need to reduce inter-node communication




                                                                              7 / 29
Introduction       Thesis    Problem    Proposed Approaches   Prior Work   Timeline




Observations


         1     We need to reduce inter-node communication

         2     Each node is learning on a data subset




                                                                               8 / 29
Introduction       Thesis    Problem    Proposed Approaches   Prior Work   Timeline




Observations


         1     We need to reduce inter-node communication

         2     Each node is learning on a data subset

         3     Can we leverage the information already learned by
               individual nodes?




                                                                               9 / 29
Introduction       Thesis    Problem    Proposed Approaches   Prior Work   Timeline




Observations


         1     We need to reduce inter-node communication

         2     Each node is learning on a data subset

         3     Can we leverage the information already learned by
               individual nodes?



      Our Idea
      Use Active Learning to exchange only informative data points


                                                                              10 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                              11 / 29
Introduction   Thesis     Problem     Proposed Approaches   Prior Work   Timeline




Thesis




      Thesis Statement
      This thesis considers a distributed learning setting to learn at
      scale and aims to propose active sampling strategies that
      reduce communication overhead in such distributed setups.




                                                                            12 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                              13 / 29
Introduction        Thesis     Problem     Proposed Approaches   Prior Work   Timeline




Assumptions



         1     Noiseless

         2     Two-party Communication

         3     Distribution free setting

         4     Adversarial Partition

         5     {Instance,Label}-based communication



                                                                                 14 / 29
Introduction         Thesis          Problem         Proposed Approaches          Prior Work      Timeline




Problem

      Setup

          A                                           K        B
          DA = labeled datasets for node A                     DB = labeled datasets for node B
                                                    hAB
          hA = hypothesis learned on DA                        hB = hypothesis learned on DB

                              C
                              D = DA โˆช DB = labeled datasets for node C
                              hC = hypothesis learned on D


      Goal
                   minimize K,               such that,   er (hAB ) โ‰ค er (hC ) +
                 where K = number of communications and is constant (0 <              < 1).

                                                                                                     15 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                              16 / 29
Introduction   Thesis   Problem   Proposed Approaches   Prior Work   Timeline




Proposed Approaches




                                                                        17 / 29
Introduction       Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Extensions




         1     Agnostic

         2     Privacy

         3     Multiparty

         4     Alternate forms of communication




                                                                             18 / 29
Introduction       Thesis    Problem   Proposed Approaches   Prior Work   Timeline




My Focus



         1       Agnostic



         2       Privacy


         3     Multiparty

         4     Alternate forms of communication



                                                                             19 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                              20 / 29
Introduction        Thesis   Problem   Proposed Approaches   Prior Work   Timeline




Prior Work

      Past Projects

               Databases
                1   Metric Functional Dependencies (MFD)
                2   Sequential Dependencies (SD)


               Machine Learning
                1   Semi-supervised Domain Adaptation (EA++)
                2   Online MultiTask Learning (OMTL)
                3   Active Learning for Domain Adaptation (ALDA)


                                                                             21 / 29
Introduction        Thesis   Problem   Proposed Approaches   Prior Work   Timeline




Prior Work

      Past Projects

               Databases
                1   Metric Functional Dependencies (MFD)
                2   Sequential Dependencies (SD)


               Machine Learning
                1   Semi-supervised Domain Adaptation (EA++)
                2   Online MultiTask Learning (OMTL)
                3   Active Learning for Domain Adaptation (ALDA)


                                                                             22 / 29
Introduction      Thesis   Problem   Proposed Approaches             Prior Work              Timeline




EA++


                                     EA:                   ฯ†s
                                     EA++:                      Ls      Ls        0

         1     EA++                                        ฯ†t   Lt       0        Lt

         2     OMTL                    Ls        Lt              0      U         -U
                                                                                       R3d
         3     ALDA                                        ฯ†u
                                             U                  Ls: labeled source data
                                                                Lt: labeled target data
                                                      Rd        U : unlabeled data
         1     EA++
         2     OMTL
         3     ALDA




                                                                                                23 / 29
Introduction      Thesis   Problem   Proposed Approaches                        Prior Work    Timeline




OMTL


                                                         w            Task 1

                                                          w1
                                        A
         1     EA++                       Task
                                       Relationship
                                                          w2

                                         Matrix
         2     OMTL                                                                    Data


                                                         wK
         3     ALDA                    Omtl                           Task K


                                       Omtl = Online MultiTask Learning
                                       wi = weight vector of task i, wi โˆˆ Rd
                                       wT = (wT , wT , . . . , wT ) โˆˆ RKd
                                               1    2           K
         1     EA++                    A = Task relationship (interaction) matrix

         2     OMTL
         3     ALDA




                                                                                                 24 / 29
Introduction      Thesis   Problem   Proposed Approaches                    Prior Work        Timeline




ALDA



                                                       Target
                                                                          Active Learning
         1     EA++                    Source
                                                                       ฯ†: Domain Adaptation

                                                                                     uฯ†
         2     OMTL
         3     ALDA                  Learn classi๏ฌer
                                        on source
                                      labeled data
                                                                                     wds

                                                  Free Oracle Of
                                                       wsrc
                                     Hybrid            Costly Oracle   Learner
         1     EA++                  Oracle                 Oe         (in target)




         2     OMTL
         3     ALDA




                                                                                                 25 / 29
Introduction   Thesis   Problem   Proposed Approaches   Prior Work   Timeline




Connecting the Past and the Future




                                                                        26 / 29
Introduction        Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Outline


       1       Introduction

       2       Thesis

       3       Problem

       4       Proposed Approaches

       5       Prior Work

       6       Timeline


                                                                              27 / 29
Introduction                 Thesis              Problem                      Proposed Approaches                           Prior Work     Timeline




Timeline

          Ph.D.ย TIMELINE

        2011                                                                                                                 2012
          May         June     July    Aug       Sep        Oct        Nov       Dec        Jan          Feb     Mar       April     May


         Submitย 
         workย onย 
         noiselessย 
         setting


                  1.ย Submitย workย onย agnosticย setting
                  2.ย Lookย intoย privacyย relatedย issues


                                                        1.ย Identifyย interestingย directionsย inย privacyย 
                                                        2.ย Workย onย one/twoย privacyย problems
                                                        3.ย Submitย privacyย relatedย works


                                                                                                          1.ย Writeย thesis
                                                                                                          2.ย Interview
                                                                                                          3.ย Defendย andย graduate!ย 




                                                                                                                                              28 / 29
Introduction   Thesis    Problem   Proposed Approaches   Prior Work   Timeline




Questions?




                        THANK YOU !!



                                                                         29 / 29

More Related Content

Viewers also liked

Bonfire Night
Bonfire NightBonfire Night
Bonfire Night
mluisavm
ย 
hotel.jck.final.7.15
hotel.jck.final.7.15hotel.jck.final.7.15
hotel.jck.final.7.15
Michele Meyer
ย 
Cumhuriyet ' e nasฤฑl kavuลŸtuk
Cumhuriyet ' e nasฤฑl kavuลŸtukCumhuriyet ' e nasฤฑl kavuลŸtuk
Cumhuriyet ' e nasฤฑl kavuลŸtuksudecavdar
ย 
Michael Dell case
Michael Dell caseMichael Dell case
Michael Dell case
Pritom Kaysar
ย 
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
Adre Schreuder
ย 
General presentation sales spanish pl
General presentation sales spanish plGeneral presentation sales spanish pl
General presentation sales spanish pl
PrintLAT
ย 
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
MediaPost
ย 
CV BRIGHT 2015 March
CV BRIGHT 2015 MarchCV BRIGHT 2015 March
CV BRIGHT 2015 March
Bright Moloka
ย 
Payslip (1).Pdf
Payslip (1).PdfPayslip (1).Pdf
Payslip (1).Pdf
Er. VIVEK CHANDRA MISHRA
ย 
PEPSICO appointment letter
PEPSICO appointment letter PEPSICO appointment letter
PEPSICO appointment letter
Muhammad usman javed
ย 

Viewers also liked (10)

Bonfire Night
Bonfire NightBonfire Night
Bonfire Night
ย 
hotel.jck.final.7.15
hotel.jck.final.7.15hotel.jck.final.7.15
hotel.jck.final.7.15
ย 
Cumhuriyet ' e nasฤฑl kavuลŸtuk
Cumhuriyet ' e nasฤฑl kavuลŸtukCumhuriyet ' e nasฤฑl kavuลŸtuk
Cumhuriyet ' e nasฤฑl kavuลŸtuk
ย 
Michael Dell case
Michael Dell caseMichael Dell case
Michael Dell case
ย 
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
CXPA Networking event Jhb May 2016 - Overview of Customer Experience in South...
ย 
General presentation sales spanish pl
General presentation sales spanish plGeneral presentation sales spanish pl
General presentation sales spanish pl
ย 
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
Jen Yomoah, Brand Manager, Edgewell - How to Use Social Influencers to Reach ...
ย 
CV BRIGHT 2015 March
CV BRIGHT 2015 MarchCV BRIGHT 2015 March
CV BRIGHT 2015 March
ย 
Payslip (1).Pdf
Payslip (1).PdfPayslip (1).Pdf
Payslip (1).Pdf
ย 
PEPSICO appointment letter
PEPSICO appointment letter PEPSICO appointment letter
PEPSICO appointment letter
ย 

Similar to Phd proposal-talk

Design theorizing ecis practise workshop 09062012
Design theorizing ecis practise workshop 09062012Design theorizing ecis practise workshop 09062012
Design theorizing ecis practise workshop 09062012
Matti Rossi
ย 
Machine Learning, LIX004M5
Machine Learning, LIX004M5Machine Learning, LIX004M5
Machine Learning, LIX004M5
butest
ย 
LTHE about PBL
LTHE about PBLLTHE about PBL
LTHE about PBL
Chrissi Nerantzi
ย 
Blockcourses
BlockcoursesBlockcourses
Blockcourses
Joanna Dunlap
ย 
On Semi-Supervised Learning and Beyond
On Semi-Supervised Learning and BeyondOn Semi-Supervised Learning and Beyond
On Semi-Supervised Learning and Beyond
Eunjeong (Lucy) Park
ย 
Alianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolioAlianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolio
sburakharper
ย 
2.7.king
2.7.king2.7.king
2.7.king
afacct
ย 
mba570.springsummer09.4
mba570.springsummer09.4mba570.springsummer09.4
mba570.springsummer09.4
Lawrence
ย 
mba570.springsummer09.5
mba570.springsummer09.5mba570.springsummer09.5
mba570.springsummer09.5
Lawrence
ย 
Large Knowledge Bases
Large Knowledge BasesLarge Knowledge Bases
Large Knowledge Bases
Bruno Paiva Lima da Silva
ย 
Combatting survey fatigue: uncovering the student learning experience using the
Combatting survey fatigue: uncovering the student learning experience using the Combatting survey fatigue: uncovering the student learning experience using the
Combatting survey fatigue: uncovering the student learning experience using the
Tรผnde Varga-Atkins
ย 
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
Engineers Australia
ย 

Similar to Phd proposal-talk (12)

Design theorizing ecis practise workshop 09062012
Design theorizing ecis practise workshop 09062012Design theorizing ecis practise workshop 09062012
Design theorizing ecis practise workshop 09062012
ย 
Machine Learning, LIX004M5
Machine Learning, LIX004M5Machine Learning, LIX004M5
Machine Learning, LIX004M5
ย 
LTHE about PBL
LTHE about PBLLTHE about PBL
LTHE about PBL
ย 
Blockcourses
BlockcoursesBlockcourses
Blockcourses
ย 
On Semi-Supervised Learning and Beyond
On Semi-Supervised Learning and BeyondOn Semi-Supervised Learning and Beyond
On Semi-Supervised Learning and Beyond
ย 
Alianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolioAlianna Maren STOMP ePortfolio
Alianna Maren STOMP ePortfolio
ย 
2.7.king
2.7.king2.7.king
2.7.king
ย 
mba570.springsummer09.4
mba570.springsummer09.4mba570.springsummer09.4
mba570.springsummer09.4
ย 
mba570.springsummer09.5
mba570.springsummer09.5mba570.springsummer09.5
mba570.springsummer09.5
ย 
Large Knowledge Bases
Large Knowledge BasesLarge Knowledge Bases
Large Knowledge Bases
ย 
Combatting survey fatigue: uncovering the student learning experience using the
Combatting survey fatigue: uncovering the student learning experience using the Combatting survey fatigue: uncovering the student learning experience using the
Combatting survey fatigue: uncovering the student learning experience using the
ย 
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
ICWES15 - Today's Relevance of Feminist Theory and Gender Inclusive Engineeri...
ย 

Recently uploaded

How to Fix [Errno 98] address already in use
How to Fix [Errno 98] address already in useHow to Fix [Errno 98] address already in use
How to Fix [Errno 98] address already in use
Celine George
ย 
Standardized tool for Intelligence test.
Standardized tool for Intelligence test.Standardized tool for Intelligence test.
Standardized tool for Intelligence test.
deepaannamalai16
ย 
Information and Communication Technology in Education
Information and Communication Technology in EducationInformation and Communication Technology in Education
Information and Communication Technology in Education
MJDuyan
ย 
Ch-4 Forest Society and colonialism 2.pdf
Ch-4 Forest Society and colonialism 2.pdfCh-4 Forest Society and colonialism 2.pdf
Ch-4 Forest Society and colonialism 2.pdf
lakshayrojroj
ย 
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
indexPub
ย 
Temple of Asclepius in Thrace. Excavation results
Temple of Asclepius in Thrace. Excavation resultsTemple of Asclepius in Thrace. Excavation results
Temple of Asclepius in Thrace. Excavation results
Krassimira Luka
ย 
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptxRESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
zuzanka
ย 
A Visual Guide to 1 Samuel | A Tale of Two Hearts
A Visual Guide to 1 Samuel | A Tale of Two HeartsA Visual Guide to 1 Samuel | A Tale of Two Hearts
A Visual Guide to 1 Samuel | A Tale of Two Hearts
Steve Thomason
ย 
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGHKHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
shreyassri1208
ย 
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
ImMuslim
ย 
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) CurriculumPhilippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
MJDuyan
ย 
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
Nguyen Thanh Tu Collection
ย 
How to Download & Install Module From the Odoo App Store in Odoo 17
How to Download & Install Module From the Odoo App Store in Odoo 17How to Download & Install Module From the Odoo App Store in Odoo 17
How to Download & Install Module From the Odoo App Store in Odoo 17
Celine George
ย 
220711130097 Tulip Samanta Concept of Information and Communication Technology
220711130097 Tulip Samanta Concept of Information and Communication Technology220711130097 Tulip Samanta Concept of Information and Communication Technology
220711130097 Tulip Samanta Concept of Information and Communication Technology
Kalna College
ย 
BPSC-105 important questions for june term end exam
BPSC-105 important questions for june term end examBPSC-105 important questions for june term end exam
BPSC-105 important questions for june term end exam
sonukumargpnirsadhan
ย 
Bonku-Babus-Friend by Sathyajith Ray (9)
Bonku-Babus-Friend by Sathyajith Ray  (9)Bonku-Babus-Friend by Sathyajith Ray  (9)
Bonku-Babus-Friend by Sathyajith Ray (9)
nitinpv4ai
ย 
Level 3 NCEA - NZ: A Nation In the Making 1872 - 1900 SML.ppt
Level 3 NCEA - NZ: A  Nation In the Making 1872 - 1900 SML.pptLevel 3 NCEA - NZ: A  Nation In the Making 1872 - 1900 SML.ppt
Level 3 NCEA - NZ: A Nation In the Making 1872 - 1900 SML.ppt
Henry Hollis
ย 
Contiguity Of Various Message Forms - Rupam Chandra.pptx
Contiguity Of Various Message Forms - Rupam Chandra.pptxContiguity Of Various Message Forms - Rupam Chandra.pptx
Contiguity Of Various Message Forms - Rupam Chandra.pptx
Kalna College
ย 
220711130088 Sumi Basak Virtual University EPC 3.pptx
220711130088 Sumi Basak Virtual University EPC 3.pptx220711130088 Sumi Basak Virtual University EPC 3.pptx
220711130088 Sumi Basak Virtual University EPC 3.pptx
Kalna College
ย 
HYPERTENSION - SLIDE SHARE PRESENTATION.
HYPERTENSION - SLIDE SHARE PRESENTATION.HYPERTENSION - SLIDE SHARE PRESENTATION.
HYPERTENSION - SLIDE SHARE PRESENTATION.
deepaannamalai16
ย 

Recently uploaded (20)

How to Fix [Errno 98] address already in use
How to Fix [Errno 98] address already in useHow to Fix [Errno 98] address already in use
How to Fix [Errno 98] address already in use
ย 
Standardized tool for Intelligence test.
Standardized tool for Intelligence test.Standardized tool for Intelligence test.
Standardized tool for Intelligence test.
ย 
Information and Communication Technology in Education
Information and Communication Technology in EducationInformation and Communication Technology in Education
Information and Communication Technology in Education
ย 
Ch-4 Forest Society and colonialism 2.pdf
Ch-4 Forest Society and colonialism 2.pdfCh-4 Forest Society and colonialism 2.pdf
Ch-4 Forest Society and colonialism 2.pdf
ย 
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
THE SACRIFICE HOW PRO-PALESTINE PROTESTS STUDENTS ARE SACRIFICING TO CHANGE T...
ย 
Temple of Asclepius in Thrace. Excavation results
Temple of Asclepius in Thrace. Excavation resultsTemple of Asclepius in Thrace. Excavation results
Temple of Asclepius in Thrace. Excavation results
ย 
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptxRESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
RESULTS OF THE EVALUATION QUESTIONNAIRE.pptx
ย 
A Visual Guide to 1 Samuel | A Tale of Two Hearts
A Visual Guide to 1 Samuel | A Tale of Two HeartsA Visual Guide to 1 Samuel | A Tale of Two Hearts
A Visual Guide to 1 Samuel | A Tale of Two Hearts
ย 
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGHKHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
KHUSWANT SINGH.pptx ALL YOU NEED TO KNOW ABOUT KHUSHWANT SINGH
ย 
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
Geography as a Discipline Chapter 1 __ Class 11 Geography NCERT _ Class Notes...
ย 
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) CurriculumPhilippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
Philippine Edukasyong Pantahanan at Pangkabuhayan (EPP) Curriculum
ย 
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
CHUYรŠN ฤแป€ ร”N TแบฌP Vร€ PHรT TRIแป‚N Cร‚U HแปŽI TRONG ฤแป€ MINH HแปŒA THI TแปT NGHIแป†P THPT ...
ย 
How to Download & Install Module From the Odoo App Store in Odoo 17
How to Download & Install Module From the Odoo App Store in Odoo 17How to Download & Install Module From the Odoo App Store in Odoo 17
How to Download & Install Module From the Odoo App Store in Odoo 17
ย 
220711130097 Tulip Samanta Concept of Information and Communication Technology
220711130097 Tulip Samanta Concept of Information and Communication Technology220711130097 Tulip Samanta Concept of Information and Communication Technology
220711130097 Tulip Samanta Concept of Information and Communication Technology
ย 
BPSC-105 important questions for june term end exam
BPSC-105 important questions for june term end examBPSC-105 important questions for june term end exam
BPSC-105 important questions for june term end exam
ย 
Bonku-Babus-Friend by Sathyajith Ray (9)
Bonku-Babus-Friend by Sathyajith Ray  (9)Bonku-Babus-Friend by Sathyajith Ray  (9)
Bonku-Babus-Friend by Sathyajith Ray (9)
ย 
Level 3 NCEA - NZ: A Nation In the Making 1872 - 1900 SML.ppt
Level 3 NCEA - NZ: A  Nation In the Making 1872 - 1900 SML.pptLevel 3 NCEA - NZ: A  Nation In the Making 1872 - 1900 SML.ppt
Level 3 NCEA - NZ: A Nation In the Making 1872 - 1900 SML.ppt
ย 
Contiguity Of Various Message Forms - Rupam Chandra.pptx
Contiguity Of Various Message Forms - Rupam Chandra.pptxContiguity Of Various Message Forms - Rupam Chandra.pptx
Contiguity Of Various Message Forms - Rupam Chandra.pptx
ย 
220711130088 Sumi Basak Virtual University EPC 3.pptx
220711130088 Sumi Basak Virtual University EPC 3.pptx220711130088 Sumi Basak Virtual University EPC 3.pptx
220711130088 Sumi Basak Virtual University EPC 3.pptx
ย 
HYPERTENSION - SLIDE SHARE PRESENTATION.
HYPERTENSION - SLIDE SHARE PRESENTATION.HYPERTENSION - SLIDE SHARE PRESENTATION.
HYPERTENSION - SLIDE SHARE PRESENTATION.
ย 

Phd proposal-talk

  • 1. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Active Communication for Distributed Learning Avishek Saha School Of Computing, University of Utah PhD Proposal 4th May, 2011 1 / 29
  • 2. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 2 / 29
  • 3. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 3 / 29
  • 4. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Large scale learning 1 Machine learning on large scale data 2 Data size surpasses the limits of a single core 3 Distributed learning on multiple cores 4 / 29
  • 5. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Challenges 1 Parallelization of serial algorithms/new parallel algorithms 2 Scheduling issues in multiple cores 3 Combining results from multiple cores 4 Communication bottlenecks in distributed learning 5 / 29
  • 6. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Challenges 1 Parallelization of serial algorithms/new parallel algorithms 2 Scheduling issues in multiple cores 3 Combining results from multiple cores ยง ยค 4 Communication bottlenecks in distributed learning ยฆ ยฅ 6 / 29
  • 7. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Observations 1 We need to reduce inter-node communication 7 / 29
  • 8. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Observations 1 We need to reduce inter-node communication 2 Each node is learning on a data subset 8 / 29
  • 9. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Observations 1 We need to reduce inter-node communication 2 Each node is learning on a data subset 3 Can we leverage the information already learned by individual nodes? 9 / 29
  • 10. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Observations 1 We need to reduce inter-node communication 2 Each node is learning on a data subset 3 Can we leverage the information already learned by individual nodes? Our Idea Use Active Learning to exchange only informative data points 10 / 29
  • 11. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 11 / 29
  • 12. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Thesis Thesis Statement This thesis considers a distributed learning setting to learn at scale and aims to propose active sampling strategies that reduce communication overhead in such distributed setups. 12 / 29
  • 13. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 13 / 29
  • 14. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Assumptions 1 Noiseless 2 Two-party Communication 3 Distribution free setting 4 Adversarial Partition 5 {Instance,Label}-based communication 14 / 29
  • 15. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Problem Setup A K B DA = labeled datasets for node A DB = labeled datasets for node B hAB hA = hypothesis learned on DA hB = hypothesis learned on DB C D = DA โˆช DB = labeled datasets for node C hC = hypothesis learned on D Goal minimize K, such that, er (hAB ) โ‰ค er (hC ) + where K = number of communications and is constant (0 < < 1). 15 / 29
  • 16. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 16 / 29
  • 17. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Proposed Approaches 17 / 29
  • 18. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Extensions 1 Agnostic 2 Privacy 3 Multiparty 4 Alternate forms of communication 18 / 29
  • 19. Introduction Thesis Problem Proposed Approaches Prior Work Timeline My Focus 1 Agnostic 2 Privacy 3 Multiparty 4 Alternate forms of communication 19 / 29
  • 20. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 20 / 29
  • 21. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Prior Work Past Projects Databases 1 Metric Functional Dependencies (MFD) 2 Sequential Dependencies (SD) Machine Learning 1 Semi-supervised Domain Adaptation (EA++) 2 Online MultiTask Learning (OMTL) 3 Active Learning for Domain Adaptation (ALDA) 21 / 29
  • 22. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Prior Work Past Projects Databases 1 Metric Functional Dependencies (MFD) 2 Sequential Dependencies (SD) Machine Learning 1 Semi-supervised Domain Adaptation (EA++) 2 Online MultiTask Learning (OMTL) 3 Active Learning for Domain Adaptation (ALDA) 22 / 29
  • 23. Introduction Thesis Problem Proposed Approaches Prior Work Timeline EA++ EA: ฯ†s EA++: Ls Ls 0 1 EA++ ฯ†t Lt 0 Lt 2 OMTL Ls Lt 0 U -U R3d 3 ALDA ฯ†u U Ls: labeled source data Lt: labeled target data Rd U : unlabeled data 1 EA++ 2 OMTL 3 ALDA 23 / 29
  • 24. Introduction Thesis Problem Proposed Approaches Prior Work Timeline OMTL w Task 1 w1 A 1 EA++ Task Relationship w2 Matrix 2 OMTL Data wK 3 ALDA Omtl Task K Omtl = Online MultiTask Learning wi = weight vector of task i, wi โˆˆ Rd wT = (wT , wT , . . . , wT ) โˆˆ RKd 1 2 K 1 EA++ A = Task relationship (interaction) matrix 2 OMTL 3 ALDA 24 / 29
  • 25. Introduction Thesis Problem Proposed Approaches Prior Work Timeline ALDA Target Active Learning 1 EA++ Source ฯ†: Domain Adaptation uฯ† 2 OMTL 3 ALDA Learn classi๏ฌer on source labeled data wds Free Oracle Of wsrc Hybrid Costly Oracle Learner 1 EA++ Oracle Oe (in target) 2 OMTL 3 ALDA 25 / 29
  • 26. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Connecting the Past and the Future 26 / 29
  • 27. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Outline 1 Introduction 2 Thesis 3 Problem 4 Proposed Approaches 5 Prior Work 6 Timeline 27 / 29
  • 28. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Timeline Ph.D.ย TIMELINE 2011 2012 May June July Aug Sep Oct Nov Dec Jan Feb Mar April May Submitย  workย onย  noiselessย  setting 1.ย Submitย workย onย agnosticย setting 2.ย Lookย intoย privacyย relatedย issues 1.ย Identifyย interestingย directionsย inย privacyย  2.ย Workย onย one/twoย privacyย problems 3.ย Submitย privacyย relatedย works 1.ย Writeย thesis 2.ย Interview 3.ย Defendย andย graduate!ย  28 / 29
  • 29. Introduction Thesis Problem Proposed Approaches Prior Work Timeline Questions? THANK YOU !! 29 / 29