Deep Learning vs. Cheap Learning

© 2016 MapR Technologies 3
Me, Us
• Ted Dunning, MapR Chief Application Architect, Apache Member
– Committer PMC member Zookeeper, Drill, others
– Mentor for Flink, Beam (nee Dataflow), Drill, Storm, Zeppelin
– VP Incubator
– Bought the beer at the first HUG
• MapR
– Produces first converged platform for big and fast data
– Includes data platform (files, streams, tables) + open source
– Adds major technology for performance, HA, industry standard API’s
• Contact
@ted_dunning, ted.dunning@gmail.com, tdunning@mapr.com

Agenda
• Rationale
• Why cheap isn't the same as simple-minded
• Some techniques
• Examples

Outline
• We have a revolution on our hands
• This leads to a green-field situation
• That implies that many important problems are easy to solve
• The limiting factor is fielding good enough solutions
– Quickly
– With available workforce
• Examples

Is this really a
revolutionary moment?

Big is the next big thing
• Data scale is exploding
• Companies are being funded
• Books are being written
• Applications sprouting up everywhere

Why Now?
• But Moore’s law has applied for a long time
• Why is data exploding now?
• Why not 10 years ago?
• Why not 20?

Size Matters, but …
• If it were just availability of data then existing big companies
would adopt big data technology first

Size Matters, but …
• If it were just availability of data then existing big companies
would adopt big data technology first
They didn’t

Or Maybe Cost
• If it were just a net positive value then finance companies should
adopt first because they have higher opportunity value / byte

Or Maybe Cost
• If it were just a net positive value then finance companies should
adopt first because they have higher opportunity value / byte
They didn’t

Backwards adoption
• Under almost any threshold argument startups would not adopt
big data technology first

Backwards adoption
• Under almost any threshold argument startups would not adopt
big data technology first
They did

Everywhere at Once?
• Something very strange is happening
– Big data is being applied at many different scales
– At many value scales
– By large companies and small

Everywhere at Once?
• Something very strange is happening
– Big data is being applied at many different scales
– At many value scales
– By large companies and small
Why?

Analytics Scaling Laws
• Analytics scaling is all about the 80-20 rule
– Big gains for little initial effort
– Rapidly diminishing returns
• The key to net value is how costs scale
– Old school – exponential scaling
– Big data – linear scaling, low constant
• Cost/performance has changed radically
– IF you can use many commodity boxes

Most data isn’t worth much in isolation
First data is valuable
Later data is dregs

Suddenly worth processing
First data is valuable
Later data is dregs
But has high aggregate value

If we can handle the scale
It’s really big

So what makes
that possible?

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value
2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value
Net value optimum has
a sharp peak well
before maximum effort

But scaling laws are
changing both slope and
shape

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value
More than just a
little

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value
They are changing a
LOT!

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value

2,0000 500 1000 1500
1
0
0.25
0.5
0.75
Scale
Value
Initially, linear cost scaling
actually makes things
worse
Then a tipping point is
reached and things change
radically …

Pre-requisites for Tipping
• To reach the tipping point,
• Algorithms must scale out horizontally
– On commodity hardware
– That can and will fail
• Data practice must change
– Denormalized is the new black
– Flexible data dictionaries are the rule
– Structured data becomes rare

With great scale comes great opportunity
• Increasing scale by 1000x changes the game
• We essentially have green fields opening up all around
• Most of the opportunities don’t require advanced learning

OK.
We have a
bona fide revolution

Greenfield Problem Landscape

Mature Problem Landscape

Why is cheap better than deep (sometimes)?
When we have a greenfield, problems can be
– Easy (large number of these)
– Impossible (large number of these)
– Hard but possible (right on the boundary)
In a mature field, problems can be
– Easy (these are already done)
– Impossible (still a large number of these)
– Hard but possible (now the majority of the effort)

Some examples

A simple example - security monitoring
• “Small” data
– Capture IDS logs
– Detect what you already know
• “Big” data
– Capture switch, server, firewall logs as well
– New patterns emerge immediately

Another example – fraud detection
• “Small” data
– Maintain card profiles
– Segment models
– Evaluate all transactions
• “Big” Data
– Maintain card profiles, full 90 day transaction history
– Evaluate all transactions

Another example – indicator-based recommendation
• “Advanced” approach
– Use matrix completion techniques (LDA, NNM, ALS)
– Tune meta-parameters
– Ensembles galore
• “Simple” approach
– Count cooccurrences and cross-occurrences
– Finding “interesting” pairs
– Use standard search engine to recommend

Easy != Stupid
• You still have to do things reasonably well
– Techniques that are not well founded are still problems
• Heuristic frequency ratios still fail
– Coincidences still dominate the data
– Accidental 100% correlations abound
• Related techniques still broken for coincidence
– Pearson’s χ2
– Simple correlations

Scale does not cure wrong
It just makes easy more common

A core technique
• Many of these easy problems reduce to finding interesting
coincidences
• This can be summarized as a 2 x 2 table
• Actually, many of these tables
A Other
B k11 k12
Other k21 k22

How do you do that?
• This is well handled using G2-test
– See wikipedia
– See http://bit.ly/surprise-and-coincidence
• Original application in linguistics now cited > 2000 times
• Available in ElasticSearch, in Solr, in Mahout
• Available in R, C, Java, Python

Which one is the anomalous co-occurrence?
A not A
B 13 1000
not B 1000 100,000
A not A
B 1 0
not B 0 10,000
A not A
B 10 0
not B 0 100,000
A not A
B 1 0
not B 0 2

Which one is the anomalous co-occurrence?
A not A
B 13 1000
not B 1000 100,000
A not A
B 1 0
not B 0 10,000
A not A
B 10 0
not B 0 100,000
A not A
B 1 0
not B 0 2
0.90 1.95
4.52 14.3
Dunning Ted, Accurate Methods for the Statistics of Surprise and Coincidence,
Computational Linguistics vol 19 no. 1 (1993)

So we can find
interesting coincidences.
That gets us exactly what?

Operation Ababil – Brobots on Parade
• Dork attack to find unpatched default Joomla sites
– Especially web servers with high bandwidth connections
– Basically just Google searches for default strings
– Joomla compromised into attack Brobot
• C&C network checks in occasionally
– Note C&C is incoming request and looks like normal web requests
• Later, on command, multiple Brobots direct 50-75 Gb/s of attack
– Attacks come from white-listed sites

Attack Sequence
Source
First level
C&C
Second
level C&C

Google
Attack Sequence
Source
First level
C&C
Second
level C&C

Brobot
Brobot
Brobot
Attack Sequence
Source
First level
C&C
Second
level C&C

Target
Brobot
Brobot
Brobot
Attack Sequence
Source
First level
C&C
Second
level C&C

Outline of an Advanced Persistent Threat
• Advanced
– Common use of zero-day for preliminary attacks
– Often attributed to state-level actors
– Modern privateers blur the line
• Persistent
– Result of first attack is heavily muffled, no immediate exploit
– Remote access toolset installed (RAT)
• Threat
– On command, data is exfiltrated covertly or en masse
– Or the compromised host is used for other nefarious purpose

APT in Summary
• Attack, penetrate, pivot, exfiltrate or exploit
• If you are a high-value target, attack is likely and stealthy
– High-value = telecom, banks, utilities, retail targets, web100
– … and all their vendors
– Conventional multi-factor auth is easily breached
• Penetration and pivot are critical counter-measure opportunities
– In 2010, RAT would contact command and control (C&C)
– In 2016, C&C looks like normal traffic
• Once exfiltration or exploit starts, you may no longer have a
business

Target
Brobot
Brobot
Brobot
Example 1 - Ababil
Source
First level
C&C
Second
level C&C
Defense has to
happen here

Spot the Important Difference?
GET /personal/comparison-table.jsp?iODg2OQ=51a90 HTTP/1.1
Host: www.sometarget.com
User-Agent: Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1;)
Accept-Encoding: deflate
Accept-Charset: UTF-8
Accept-Language: fr
Cache-Control: no-cache
Pragma: no-cache
Connection: Keep-Alive
GET /photo.jpg HTTP/1.1
Host: lh4.googleusercontent.com
User-Agent: Mozilla/5.0 (Macint
Accept: image/png,image/*;q=0.8
Accept-Language: en-US,en;q=0.5
Accept-Encoding: gzip, deflate,
Referer: https://www.google.com
Connection: keep-alive
If-None-Match: "v9”
Cache-Control: max-age=0
Attacker request Real request

This could only be found at scale

This could only be found at scale
But at scale, it is stupidly simple
to find

Target
Brobot
Brobot
Brobot
Overall Outline Again
Source
First level
C&C
Second
level C&C
Tradecraft error!

Large corpus analysis of source
IP’s wins big

Example 2 - Common Point of Compromise
• Scenario:
– Merchant 0 is compromised, leaks account data during compromise
– Fraud committed elsewhere during exploit
– High background level of fraud
– Limited detection rate for exploits
• Goal:
– Find merchant 0
• Meta-goal:
– Screen algorithms for this task without leaking sensitive data

Example 2 - Common Point of Compromise
skim exploit
Merchant 0
Skimmed
data
Merchant n
Card data is stolen
from Merchant 0
That data is used
in frauds at other
merchants

Simulation Setup
0 20 40 60 80 100
0100300500
day
count
Compromise period
Exploit period
compromises
frauds

Detection Strategy
• Select histories that precede non-fraud
• And histories that precede fraud detection
• Analyze 2x2 cooccurrence of merchant n versus fraud detection

What about the
real world?

●●●●●●●●●●●●●●●●●●●● ● ●● ●●● ●●● ●●●●● ●●●●● ●●● ●●● ●● ● ●● ●● ●● ● ●●●● ●●●● ●● ●●●● ●●●● ●●● ●● ●● ● ●● ● ●●●● ●● ● ●●●● ●●●●●● ●● ●● ●●● ●●● ●●●●● ● ●●● ●● ●●● ●●● ●● ●●●● ●
●● ●●● ●●● ●
●
● ●●
●
●
●
●●
020406080
LLR score for real data
Number of Merchants
BreachScore(LLR)
Real truly bad guys
100
101
102
103
104
105
106
Really truly bad guys

Historical cooccurrence gives high
S/N

Historical cooccurrence gives high
S/N
(we win)

Cooccurrence AnalysisCooccurrence Analysis

Real-life example
• Query: “Paco de Lucia”
• Conventional meta-data search results:
– “hombres de paco” times 400
– not much else
• Recommendation based search:
– Flamenco guitar and dancers
– Spanish and classical guitar
– Van Halen doing a classical/flamenco riff

Real-life example

So …
• There are suddenly lots of these problems
• Simple techniques have surprising power at scale
– Cooccurrence via G2 / LLR
– Distributional anomaly detection via t-digest
• These simple techniques are largely unsuitable for academic
research
• But they are highly applicable in resource constrained industrial
settings

●●●●●●●●●●●●●●●●●●●● ● ●● ●●● ●●● ●●●●● ●●●●● ●●● ●●● ●● ● ●● ●● ●● ● ●●●● ●●●● ●● ●●●● ●●●● ●●● ●● ●● ● ●● ● ●●●● ●● ● ●●●● ●●●●●● ●● ●● ●●● ●●● ●●●●● ● ●●● ●● ●●● ●●● ●● ●●●● ●
●● ●●● ●●● ●
●
● ●●
●
●
●
●●
020406080
LLR score for real data
Number of Merchants
BreachScore(LLR)
Real truly bad guys
100
101
102
103
104
105
106
Cooccurrence An
Summary
• We live in a golden age of newly achieved scale
• That scale has lowered the tree
– Hard problems are much easier
– Lots of low-hanging fruit all around us
• Cheap learning has huge value
• Code available at
http://github.com/tdunning 0 20 40 60 80 100
0100300500
day
count
Compromise period
Exploit period
compromises
frauds

Me, Us
• Ted Dunning, MapR Chief Application Architect, Apache Member
– Committer PMC member Zookeeper, Drill, others
– Mentor for Flink, Beam (nee Dataflow), Drill, Storm, Zeppelin
– VP Incubator
– Bought the beer at the first HUG
• MapR
– Produces a converged platform for big and fast data
– Includes data platform (files, streams, tables) + open source
– Adds major technology for performance, HA, industry standard API’s
• Contact
@ted_dunning, ted.dunning@gmail.com, tdunning@mapr.com

Q & A

Deep Learning vs. Cheap Learning

More Related Content

What's hot (20)

Viewers also liked (14)

Similar to Deep Learning vs. Cheap Learning (20)

More from MapR Technologies (20)

Recently uploaded (20)

Deep Learning vs. Cheap Learning

Editor's Notes