[PDF] - 16-11-04 Statistical Science and Data Science Nancy Reid 27 PDF Document

SLIDE 1

16-11-04 1

Fisher Memorial Lecture 27 Oct 2016

Statistical Science and Data Science

Nancy Reid 27 October 2016

Fisher Memorial Lecture 27 Oct 2016 2

Fisher Number

Fisher Memorial Lecture 27 Oct 2016 3

Selected Correspondence of

R. A. Fisher

Edited by J.H. Bennett

Fisher Number

Fisher Memorial Lecture 27 Oct 2016 4

“Do not forget to look up Walter Bodmer, who also has some experience being ‘bawled down’ by the Neymanians” 11 Jan 1962

SLIDE 2

16-11-04 2

“Some aspect of big data”

= Big Machines = Lots of Computing = Complex Architectures = Computer Science

Fisher Memorial Lecture 27 Oct 2016 5

Small data

= equations and formulas = mathematical modelling = a little computing = Statistical Science

Fisher Memorial Lecture 27 Oct 2016 6

p(v, h; η) ∝ 1 Z(η) exp{aT v + bT h + vT Wh}, η = (a, b, W)

Big Data

Interesting
Detailed
Informative
Fun

Fisher Memorial Lecture 27 Oct 2016 7

Small Data

Fisher Memorial Lecture 27 Oct 2016 8

So yesterday

SLIDE 3

16-11-04 3

Small Data

Fisher Memorial Lecture 27 Oct 2016 9

Big Data 2013

Fisher Memorial Lecture 27 Oct 2016 11

Gartner Hype Cycle

Fisher Memorial Lecture 27 Oct 2016 12

Big Data 2014

SLIDE 4

16-11-04 4

Fisher Memorial Lecture 27 Oct 2016 13

2015

Machine Learning

The push back

Fisher Memorial Lecture 27 Oct 2016 14

The push back

Fisher Memorial Lecture 27 Oct 2016 15

“Big data” has arrived, but big insights have not

The push back

Fisher Memorial Lecture 27 Oct 2016 16

How big data threatens democracy and increases inequality

“if the assessment never asks about race, how could the algorithm throw up racially biased results?” “Credit scores are used by nearly half of American employers to screen potential employees”

SLIDE 5

16-11-04 5

Canadian Institute for Statistical Sciences

Pacific Institute for Mathematical Sciences Centre de Recherches Mathématiques Fields Institute for Resesarch in the Mathematical Sciences

Workshops

Opening Conference and Bootcamp
Statistical Machine Learning
Optimization and Matrix Methods
Visualization: Strategies and Principles
Big Data in Health Policy
Big Data for Social Policy
Networks, Web mining, and Cyber-security
Statistical Theory for Large-scale Data
Challenges in Environmental Science
Complex Spatio-temporal Data
Commercial and Retail Banking

Fisher Memorial Lecture 27 Oct 2016 19

Opening Conference and Bootcamp

Introduction to topics at following workshops One day on each topic Many speakers started by trying to define big data “I shall not today attempt further to define the kinds of material I understand to be embraced within that shorthand description, and perhaps I could never succeed in intelligibly doing so. But I know it when I see it … ”

Justice Potter Stewart; Jacobellis v. Ohio 22 June 1964 Robert Bell, Google, Plenary Opening Lecture

Fisher Memorial Lecture 27 Oct 2016 20

SLIDE 6

16-11-04 6

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy

Fisher Memorial Lecture 27 Oct 2016 21

Some highlights

Statistical Machine Learning

Fisher Memorial Lecture 27 Oct 2016 22

Statistical Machine Learning

Fisher Memorial Lecture 27 Oct 2016 23

η = (a, b, W)

f(v, h; η) ∝ 1 Z(η) exp{aT v + bT h + vT Wh}

Restricted Boltzmann machine

Fisher Memorial Lecture 27 Oct 2016 24

natural gradient ascent
uses Fisher information as metric tensor

Girolami and Calderhead (2011); Amari (1987); Rao (1945)

Gaussian graphical model approximation to force

sparse inverse

Grosse and Salakhutdinov (2016) 32nd Internat. Conf. on Machine Learning

⌘ ⌘ + ✏ i(⌘)−1rη`(⌘; v, h) i = E(−`00)

f(v, h; η) ∝ 1 Z(η) exp{aT v + bT h + vT Wh}

` = log f

SLIDE 7

16-11-04 7

Restricted Boltzmann machine

Fisher Memorial Lecture 27 Oct 2016 25

if just one binary top node, model for

is a logistic regression

with several binary top nodes, model for

is also a logistic regression, with odds ratio depending

nly on
deep learning has ~10 layers, with millions of units

in each layer

estimating parameters is an optimization problem

h | v ht | v, h−t v

f(v, h; η) ∝ 1 Z(η) exp{aT v + bT h + vT Wh}

Restricted Boltzmann machine

Fisher Memorial Lecture 27 Oct 2016 26

Leung et al Bioinformatics 2014

Brendan Frey, Infinite Genomes Project

FieldsLive January 27 2015

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy

Fisher Memorial Lecture 27 Oct 2016 27

Some highlights

Optimization

Fisher Memorial Lecture 27 Oct 2016 28

max

θ { 1

n

X

i=1

log f(yi | xi; θ) − Pλ(θ)}

SLIDE 8

16-11-04 8

Optimization

Fisher Memorial Lecture 27 Oct 2016 29

max

θ { 1

n

X

i=1

log f(yi | xi; θ) − Pλ(θ)}

lasso penalty
is convex relaxation of
many interesting penalties are non-convex
optimization routines may not find global optimum

||θ||0

Pλ(θ) = λ||θ||1 = λΣ|θj| ||θ||1

Optimization

Fisher Memorial Lecture 27 Oct 2016 30

max

θ { 1

n

X

i=1

log f(yi | xi; θ) − Pλ(θ)}

statistical error neighbourhood of true value
approximation error iterating over t

ˆ θ − θ∗ θt − ˆ θ

Wainwright FieldsLive Jan 16 2015 Loh and Wainwright JMLR 2015

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy

Fisher Memorial Lecture 27 Oct 2016 31

Some highlights

Visualization

Fisher Memorial Lecture 27 Oct 2016 32

Innovis.cpsc.ucalgary.ca

SLIDE 9

16-11-04 9

Visualization

statistical graphics

– data representation – data exploration – filtering, sampling aggregation

information visualization
scientific visualization
cognitive science and design

Fisher Memorial Lecture 27 Oct 2016 33

Visualization

Fisher Memorial Lecture 27 Oct 2016 34

KPMG Data Observatory, IC

Visualization

Fisher Memorial Lecture 27 Oct 2016 35

KPMG Data Observatory, IC

Visualization

Fisher Memorial Lecture 27 Oct 2016 36

fivethirtyeight.com

SLIDE 10

16-11-04 10

Visualization

Fisher Memorial Lecture 27 Oct 2016 37

fivethirtyeight.com

Visualization

Fisher Memorial Lecture 27 Oct 2016 38

New York Times “The duty of beauty”

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy

Fisher Memorial Lecture 27 Oct 2016 39

Some highlights

Health Policy

Fisher Memorial Lecture 27 Oct 2016 40

SLIDE 11

16-11-04 11

Health Policy Administrative Databases

Fisher Memorial Lecture 27 Oct 2016 41

Institute for Clinical and Evaluative Sciences

Health Policy Administrative Databases

Fisher Memorial Lecture 27 Oct 2016 42

Institute for Clinical and Evaluative Science Thérèse Stukel, ICES

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy

Fisher Memorial Lecture 27 Oct 2016 44

SLIDE 12

16-11-04 12

Some highlights

Social Policy

Fisher Memorial Lecture 27 Oct 2016 45

Thérèse Stukel, ICES

Privacy

“Big Data and Innovation, Setting the Record Straight:

De-identification Does Work”

Privacy Commissioner of Ontario, July 2014

“No silver bullet: De-identification still doesn’t work”

Narayan & Felten, July 2014

Statistical Disclosure Limitation
Differential Privacy
Multi-party Communication

Fisher Memorial Lecture 27 Oct 2016 47

Some highlights

Statistical Machine Learning
Optimization
Visualization
Health Policy
Social Policy
inference, environmental science, networks, genomics,

finance, physical sciences, software infrastructure, …

Fisher Memorial Lecture 27 Oct 2016 48

SLIDE 13

16-11-04 13

What did we learn?

Statistical models for big data are complex,

high-dimensional

– inference is well-studied, but difficult

Computational challenges include size and speed

– ideas of statistical inference get lost in the machine

Data owners understand 2., but not 1.
Data science may be the best way to combine these

Fisher Memorial Lecture 27 Oct 2016 49

What is data science?

a course?
a set of courses?
a job?
a technology?
a new field of research?
a collaboration?

Fisher Memorial Lecture 27 Oct 2016 50

Data Science Program(s)

mathematical reasoning
statistical theory
statistical and machine learning methods
programming and software development
algorithms and data structure
communication results and limitations

Fisher Memorial Lecture 27 Oct 2016 51

Data Science Research

data collection and data quality
large N, small p

– computational strategies, e.g. Spark, Hadoop – divide and conquer

small n, large p

– inferential and computational strategies – dimension reduction – post-selection inference – inference for extremes

‘new’ types of data: networks, graphs, text, images, …

– “alternative sources”

Fisher Memorial Lecture 27 Oct 2016 52

SLIDE 14

16-11-04 14

… Data Science Research

collaboration and communication
data wrangling, database development, record linkage
replicability, reproducibility, new workflows
visualization
outside the ivory tower -- industry, government,

media, public

Fisher Memorial Lecture 27 Oct 2016 53 Fisher Memorial Lecture 27 Oct 2016 54

hEp://arxiv.org/abs/1609.00037v1

… Good Enough

Data Management – from raw to ‘analysable’

knitr

Software – programming

tidyr

Collaboration

dplyr

Project Organization

ggplot2

Keeping Track

ggvis

Writing

Github

Fisher Memorial Lecture 27 Oct 2016 55

“How do you see your area developing in the future?”

I suspect that the new data scientists will discover that

the old core is important

and that theoretical statisticians may be in short supply
even within statistical science we are going to need

a lot of translation

as the discipline expands it will be increasingly difficult

to be a ‘polystat’

we’ll still have lots of small data, but its analysis will be

influenced by the trend to massive data

Fisher Memorial Lecture 27 Oct 2016 56

SLIDE 15

16-11-04 15

“A range of other problems”

Fisher Memorial Lecture 27 Oct 2016 57

Michael Jordan, UC Berkeley

“while I do think of neural networks as one important tool in the toolbox, I find myself surprisingly rarely going to that tool when I’m consulting out in industry. I find that industry people are often looking to solve a range of other problems, often not involving “pattern recognition” problems”

accurate answers quickly; meaningful error bars; merge various data sources; visualize and present conclusions; diagnostics; non- stationarity; targetted experiments within databases

Caution can be a good thing

Fisher Memorial Lecture 27 Oct 2016 58

“Digital Hippocratic Oath”

Caution can be a good thing

Fisher Memorial Lecture 27 Oct 2016 59

“…from data we will get the cure for cancer as well as better hospitals; schools that adapt to children’s needs making them happier and smarter; better policing and safer homes; and of course jobs.”

Guardian 2 July 2016

Fisher Memorial Lecture 27 Oct 2016 60

Gartner Hype Cycle 2016

Smart Data Discovery

SLIDE 16

16-11-04 16

Fisher Memorial Lecture 27 Oct 2016