USING DEEP LEARNING GT 8803 // FALL 2018 // JACOB LOGAS L E C T U - - PowerPoint PPT Presentation

▶
using deep learning
SMART_READER_LITE
LIVE PREVIEW

USING DEEP LEARNING GT 8803 // FALL 2018 // JACOB LOGAS L E C T U - - PowerPoint PPT Presentation

Mar 31, 2023 338 likes •1.12k views

DATA ANALYTICS USING DEEP LEARNING GT 8803 // FALL 2018 // JACOB LOGAS L E C T U R E # 0 9 : DATA VO C A L I Z AT I O N : O P T I M I Z I N G VO I C E O U T P U T O F R E L AT I O N A L DATA TODAYS PAPER Data Vocalization:

slide-1
SLIDE 1

DATA ANALYTICS USING DEEP LEARNING

GT 8803 // FALL 2018 // JACOB LOGAS

L E C T U R E # 0 9 : DATA VO C A L I Z AT I O N : O P T I M I Z I N G VO I C E O U T P U T O F R E L AT I O N A L DATA

slide-2
SLIDE 2

GT 8803 // Fall 2018

TODAY’S PAPER

  • Data Vocalization: Optimizing Voice Output
  • f Relational Data

– New dimension to data delivery – Formalize voice output optimization problem

  • Authors: Immanuel Trummer, Jiancheng

Zhu, Mark Bryan

  • Slides based on Trummer presentation @

VLDB 2017

2

slide-3
SLIDE 3

GT 8803 // Fall 2018

TODAY’S PAPER

3

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Restaurants with Traditional American cuisine and four to five stars user average rating: Upstate. John’s. The View. Restaurants with three to four stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine. La Masseria with Italian cuisine.

slide-4
SLIDE 4

GT 8803 // Fall 2018

TODAY’S AGENDA

  • Context: Data Visualization
  • Problem Overview
  • Key Idea
  • Technical Details
  • Experiments
  • Discussion

4

slide-5
SLIDE 5

GT 8803 // Fall 2018

What is Data Visualization?

  • 1987

– NSF started “Scientific Visualization”

  • Transforms data into images

– Represent information about data

  • Tool to enable User insight into Data

– Intuitive understanding of data

5

slide-6
SLIDE 6

GT 8803 // Fall 2018

What Does Visualization Do?

  • Goals

6

This Photo by Unknown Author is licensed under CC BY-SA This Photo by Unknown Author is licensed under CC BY-SA-NC This Photo by Unknown Author is licensed under CC BY

slide-7
SLIDE 7

GT 8803 // Fall 2018

What Does Visualization Do?

  • Goals

– Explore

  • Used for data exploration

6

This Photo by Unknown Author is licensed under CC BY-SA This Photo by Unknown Author is licensed under CC BY-SA-NC This Photo by Unknown Author is licensed under CC BY

slide-8
SLIDE 8

GT 8803 // Fall 2018

What Does Visualization Do?

  • Goals

– Explore

  • Used for data exploration

– Analyze

  • Used for verification

6

This Photo by Unknown Author is licensed under CC BY-SA This Photo by Unknown Author is licensed under CC BY-SA-NC This Photo by Unknown Author is licensed under CC BY

slide-9
SLIDE 9

GT 8803 // Fall 2018

What Does Visualization Do?

  • Goals

– Explore

  • Used for data exploration

– Analyze

  • Used for verification

– Present

  • Used for Communication of Results

6

This Photo by Unknown Author is licensed under CC BY-SA This Photo by Unknown Author is licensed under CC BY-SA-NC This Photo by Unknown Author is licensed under CC BY

slide-10
SLIDE 10

GT 8803 // Fall 2018

PROBLEM OVERVIEW

7

slide-11
SLIDE 11

GT 8803 // Fall 2018

8

Visualization Cons

  • Overwhelming
  • Slowing
  • Noisy
  • Re-reading
  • Skimming

This Photo by Unknown Author is licensed under CC BY-SA

slide-12
SLIDE 12

GT 8803 // Fall 2018

Audio Presentation

  • Quick
  • Concise
  • Memorable
  • Low Cognitive

Load

9

slide-13
SLIDE 13

GT 8803 // Fall 2018

Limits of Short Term Memory

  • Impose limits on information

– Receive – Process – Remember

  • Recoding to beat bottleneck
  • Information theory

10

Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review, 63(2), 81-97. http://dx.doi.org/10.1037/h0043158

slide-14
SLIDE 14

GT 8803 // Fall 2018

Overview

  • Given input relation
  • Find time-optimal vocalization
  • Constrained by

– Precision – Output structure – Memory load

11

slide-15
SLIDE 15

GT 8803 // Fall 2018

EXAMPLE

12

slide-16
SLIDE 16

GT 8803 // Fall 2018

Naive

13

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Upstate with Traditional American cuisine and four point seven five stars user average rating. Thai Castle with Thai cuisine and three point three stars user average rating. John’s with Traditional American cuisine and four point seven stars user average rating. Paris with French cuisine and three point three stars user average rating. The View with Traditional American cuisine and four point nine stars user average rating. La Masseria with Italian cuisine and three point two stars average rating.

This Photo by Unknown Author is licensed under CC BY-ND

slide-17
SLIDE 17

GT 8803 // Fall 2018

Naive

13

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Upstate with Traditional American cuisine and four point seven five stars user average rating. Thai Castle with Thai cuisine and three point three stars user average rating. John’s with Traditional American cuisine and four point seven stars user average rating. Paris with French cuisine and three point three stars user average rating. The View with Traditional American cuisine and four point nine stars user average rating. La Masseria with Italian cuisine and three point two stars average rating.

502 Characters

This Photo by Unknown Author is licensed under CC BY-ND

slide-18
SLIDE 18

GT 8803 // Fall 2018

Naive

13

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Upstate with Traditional American cuisine and four point seven five stars user average rating. Thai Castle with Thai cuisine and three point three stars user average rating. John’s with Traditional American cuisine and four point seven stars user average rating. Paris with French cuisine and three point three stars user average rating. The View with Traditional American cuisine and four point nine stars user average rating. La Masseria with Italian cuisine and three point two stars average rating.

502 Characters

This Photo by Unknown Author is licensed under CC BY-ND

slide-19
SLIDE 19

GT 8803 // Fall 2018

Naive

13

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Upstate with Traditional American cuisine and four point seven five stars user average rating. Thai Castle with Thai cuisine and three point three stars user average rating. John’s with Traditional American cuisine and four point seven stars user average rating. Paris with French cuisine and three point three stars user average rating. The View with Traditional American cuisine and four point nine stars user average rating. La Masseria with Italian cuisine and three point two stars average rating.

502 Characters

This Photo by Unknown Author is licensed under CC BY-ND

slide-20
SLIDE 20

GT 8803 // Fall 2018

Naive

13

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Upstate with Traditional American cuisine and four point seven five stars user average rating. Thai Castle with Thai cuisine and three point three stars user average rating. John’s with Traditional American cuisine and four point seven stars user average rating. Paris with French cuisine and three point three stars user average rating. The View with Traditional American cuisine and four point nine stars user average rating. La Masseria with Italian cuisine and three point two stars average rating.

502 Characters

This Photo by Unknown Author is licensed under CC BY-ND

slide-21
SLIDE 21

GT 8803 // Fall 2018

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

slide-22
SLIDE 22

GT 8803 // Fall 2018

Contexts

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

slide-23
SLIDE 23

GT 8803 // Fall 2018

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

Scopes

slide-24
SLIDE 24

GT 8803 // Fall 2018

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

502 → 416 Characters

slide-25
SLIDE 25

GT 8803 // Fall 2018

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

502 → 416 Characters

slide-26
SLIDE 26

GT 8803 // Fall 2018

More Concise

14

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

La Masseria with Italian cuisine and three point two stars average rating. Restaurants with Traditional American cuisine: Upstate with four point seven five stars user average rating. John’s with four point seven stars user average rating. The View with four point nine stars user average rating. Restaurants with three point three stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine.

502 → 416 Characters

slide-27
SLIDE 27

GT 8803 // Fall 2018

Even More Concise

15

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Restaurants with Traditional American cuisine and four to five stars user average rating: Upstate. John’s. The View. Restaurants with three to four stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine. La Masseria with Italian cuisine.

slide-28
SLIDE 28

GT 8803 // Fall 2018

Even More Concise

15

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Restaurants with Traditional American cuisine and four to five stars user average rating: Upstate. John’s. The View. Restaurants with three to four stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine. La Masseria with Italian cuisine. 416 → 267 Characters

slide-29
SLIDE 29

GT 8803 // Fall 2018

Even More Concise

15

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 John’s Traditional American 4.7 Paris French 3.3 The View Traditional American 4.9 La Masseria Italian 3.2

Restaurants with Traditional American cuisine and four to five stars user average rating: Upstate. John’s. The View. Restaurants with three to four stars user average rating: Thai Castle with Thai cuisine. Paris with French cuisine. La Masseria with Italian cuisine. 416 → 267 Characters

slide-30
SLIDE 30

GT 8803 // Fall 2018

The Optimization Problem

  • Input: relation to vocalize
  • Search Space: sequence of scopes
  • Constraints

𝐷𝑝𝑜𝑢𝑓𝑦𝑢 𝑇𝑗𝑨𝑓 ≤ 𝑇 – Categorical value domain: 𝐸𝑝𝑛𝑏𝑗𝑜 𝑇𝑗𝑨𝑓 ≤ 𝐷 – Numerical value domains: 𝑉𝑞𝑞𝑓𝑠 𝐶𝑝𝑣𝑜𝑒 ≤ 𝑀𝑝𝑥𝑓𝑠 𝐶𝑝𝑣𝑜𝑒 ∗ 𝑋

  • Objective: Minimize speaking time

16

Memory Precision Precision

slide-31
SLIDE 31

GT 8803 // Fall 2018

The Optimization Problem

  • Input: relation to vocalize
  • Search Space: sequence of scopes
  • Constraints

𝐷𝑝𝑜𝑢𝑓𝑦𝑢 𝑇𝑗𝑨𝑓 ≤ 𝑇 – Categorical value domain: 𝐸𝑝𝑛𝑏𝑗𝑜 𝑇𝑗𝑨𝑓 ≤ 𝐷 – Numerical value domains: 𝑉𝑞𝑞𝑓𝑠 𝐶𝑝𝑣𝑜𝑒 ≤ 𝑀𝑝𝑥𝑓𝑠 𝐶𝑝𝑣𝑜𝑒 ∗ 𝑋

  • Objective: Minimize speaking time

16

Memory Precision Precision NP Hard

slide-32
SLIDE 32

GT 8803 // Fall 2018

Proof

  • Represent as vertex cover
  • One edge per row
  • One vertex per category column
  • 𝛽 if vertex incident to an edge
  • Other values are mutually different
  • Vertex cover is NP hard

17

Photo by Fschwarzentruber / CC BY

slide-33
SLIDE 33

GT 8803 // Fall 2018

ALGORITHMS

18

slide-34
SLIDE 34

GT 8803 // Fall 2018

Algorithms Overview

  • Integer Programming
  • Two-Phase Algorithm
  • Greedy Approach

19

Image used with permission of Immanuel Trummer from VLDB 2017 slide deck (http://www.itrummer.org/slides/Vocalization5.pdf)

slide-35
SLIDE 35

GT 8803 // Fall 2018

Algorithms Overview

  • Integer Programming
  • Two-Phase Algorithm
  • Greedy Approach

19

Image used with permission of Immanuel Trummer from VLDB 2017 slide deck (http://www.itrummer.org/slides/Vocalization5.pdf)

slide-36
SLIDE 36

GT 8803 // Fall 2018

Integer Linear Programming (ILP)

20 Input relation Precision constraints Memory constraints

slide-37
SLIDE 37

GT 8803 // Fall 2018

Integer Linear Programming (ILP)

20 Input relation Precision constraints Memory constraints Transformation to ILP

slide-38
SLIDE 38

GT 8803 // Fall 2018

Integer Linear Programming (ILP)

20 Input relation Precision constraints Memory constraints Transformation to ILP ILP Solver Integer Linear Program Variables Linear Constraints Linear Objective

slide-39
SLIDE 39

GT 8803 // Fall 2018

Integer Linear Programming (ILP)

20 Input relation Precision constraints Memory constraints Transformation to ILP ILP Solver Solution Integer Linear Program Variables Linear Constraints Linear Objective

slide-40
SLIDE 40

GT 8803 // Fall 2018

Integer Linear Programming (ILP)

20 Input relation Precision constraints Memory constraints Transformation to ILP ILP Solver Voice Output Solution Integer Linear Program Variables Linear Constraints Linear Objective

slide-41
SLIDE 41

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

slide-42
SLIDE 42

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

Variables

slide-43
SLIDE 43

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Needed: 1|0

Variables

slide-44
SLIDE 44

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Needed: 1|0
  • Rows: [1…n]

Variables

slide-45
SLIDE 45

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category

Variables

slide-46
SLIDE 46

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

Variables

slide-47
SLIDE 47

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

Constraints

slide-48
SLIDE 48

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

slide-49
SLIDE 49

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

Row [1…n]

slide-50
SLIDE 50

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

Row [1…n]

slide-51
SLIDE 51

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

Row [1…n] Bounded Bounded

slide-52
SLIDE 52

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

Row [1…n] Sufficiently Small Δ Sufficiently Small Δ Bounded Bounded

slide-53
SLIDE 53

GT 8803 // Fall 2018

How Do We Transform?

21

Context 1: Restaurants with… Restaurants within scope 1. … Context n: Restaurants with… Restaurants within scope n.

  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)
  • Needed: 1|0
  • Rows: [1…n]
  • Attributes: Category
  • Values: Range(a,b)

1 1

Constraints

Row [1…n] Sufficiently Small Δ Sufficiently Small Δ Bounded Bounded

Optimal Slow

slide-54
SLIDE 54

GT 8803 // Fall 2018

Two-Phase

22

Phase 1: Generate Context Candidates Phase 2: Map Rows to Candidates

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 Extract 4 to 5 stars Traditional American cuisine Traditional American cuisine and 4 to 5 stars Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 4 to 5 stars Traditional American cuisine Traditional American cuisine and 4 to 5 stars Mapping

slide-55
SLIDE 55

GT 8803 // Fall 2018

Two-Phase

22

Phase 1: Generate Context Candidates Phase 2: Map Rows to Candidates

Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 Extract 4 to 5 stars Traditional American cuisine Traditional American cuisine and 4 to 5 stars Restaurant Cuisine Rating Upstate Traditional American 4.75 Thai Castle Thai 3.3 4 to 5 stars Traditional American cuisine Traditional American cuisine and 4 to 5 stars Mapping

Fast No guarantee

slide-56
SLIDE 56

GT 8803 // Fall 2018

Phase 1: Generate

23

Figure used with permission of Immanuel Trummer from Optimizing voice-based output

  • f relational data
slide-57
SLIDE 57

GT 8803 // Fall 2018

Phase 1: Generate

  • Voice Rule

– Based on apriori rule – A context is useful iff time to say less than time saved – Time saved: The potential savings from naïve

  • Lemma 1: A specialization of a useless

context is useless

  • Lemma 2: Row cover is submodular

24

slide-58
SLIDE 58

GT 8803 // Fall 2018

Phase 2: Mapping

  • Again uses integer programming
  • Much simpler than last one
  • Add an empty context
  • New optimization goal

25

slide-59
SLIDE 59

GT 8803 // Fall 2018

Greedy

26 Find Optimal Voice Output Optimal Exponential

slide-60
SLIDE 60

GT 8803 // Fall 2018

Greedy

26 Find Optimal Context Set Find Optimal Voice Output Optimal Optimal Exponential Polynomial

slide-61
SLIDE 61

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Voice Output Optimal Optimal Exponential Polynomial ? Polynomial

slide-62
SLIDE 62

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Voice Output Optimal Optimal Exponential Polynomial ? Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative
slide-63
SLIDE 63

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Voice Output Optimal Optimal Exponential Polynomial Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative

Near-Optimal

slide-64
SLIDE 64

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Context Set Find Optimal Voice Output Optimal Polynomial Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative

Near-Optimal ? Polynomial

slide-65
SLIDE 65

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Context Set Find Optimal Voice Output Optimal Polynomial Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative

Near-Optimal ? T(assignments) Properties that hold:

  • 1. Submodular
  • 2. Non-negative

Polynomial

slide-66
SLIDE 66

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Context Set Find Optimal Voice Output Optimal Polynomial Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative

Near-Optimal T(assignments) Properties that hold:

  • 1. Submodular
  • 2. Non-negative

Polynomial Near-Optimal

slide-67
SLIDE 67

GT 8803 // Fall 2018

Greedily Construct Context Set

Greedy

26 Find Optimal Context Set Find Optimal Context Set Find Optimal Voice Output Optimal Polynomial Polynomial T({context}) Properties that hold:

  • 1. Submodular
  • 2. Monotone
  • 3. Non-negative

Near-Optimal T(assignments) Properties that hold:

  • 1. Submodular
  • 2. Non-negative

Polynomial Near-Optimal

Polynomial Optimal

slide-68
SLIDE 68

GT 8803 // Fall 2018

EXPERIMENTS

27

slide-69
SLIDE 69

GT 8803 // Fall 2018

Scope

  • Restaurants
  • Mobile Phones
  • Football Statistics
  • Laptop Models

28

slide-70
SLIDE 70

GT 8803 // Fall 2018

Configurations

  • Naïve Baseline
  • Integer Programming
  • Two-Phase Algorithm
  • Greedy Approach

29

slide-71
SLIDE 71

GT 8803 // Fall 2018

Metrics

  • User Preference

– Mechanical Turks

  • Speech Length
  • Optimization Time

30

slide-72
SLIDE 72

GT 8803 // Fall 2018

User Preference

31

Image used with permission of Immanuel Trummer from VLDB 2017 slide deck (http://www.itrummer.org/slides/Vocalization5.pdf)

slide-73
SLIDE 73

GT 8803 // Fall 2018

Speech Length

32 Precision Complexity High Low Medium Low Low Low High Medium Medium Medium Low Medium

Normalized Length # Rows

slide-74
SLIDE 74

GT 8803 // Fall 2018

Optimization Time

33 Precision Complexity High Low Medium Low Low Low High Medium Medium Medium Low Medium

Normalized Length # Rows

slide-75
SLIDE 75

GT 8803 // Fall 2018

DISCUSSION

34

slide-76
SLIDE 76

GT 8803 // Fall 2018

Strengths?

  • Iterative improvement on algorithm
  • Relevance to new device interactions
  • Takes into account cognition of users
  • Good heuristics for cognitive load

35

slide-77
SLIDE 77

GT 8803 // Fall 2018

Weaknesses or Assumptions Made?

  • Audio and Video Representations are

equivalent

  • Evaluation takes place locally
  • Google’s heuristics are applicable here
  • Interaction is one-sided
  • Poor visualization of results
  • Largely ignores natural language

36