I will examine predictive technologies in the light of the history of technology and its prediction. No matter how shiny and new, a new technology is still a technology, and there are general patterns that seem to recur. We can learn from those patterns if we pay attention to them.
In particular I will look at the challenge of predicting the impact of new technologies, talk about how they evolve, and the role that modularity, standards and interoperability play in their evolution.
I will talk more specifically about some of the particular challenges of making APIs and interfaces for predictive technologies such as machine learning, and speculate on the prospects for making machine learning a service, and more of a mature engineering discipline. In passing I will briefly demonstrate some recent machine learning work from NICTA.
2. • ML researcher & leader of NICTA ML group
– Algorithms, such as one-class SVM
– Theory – generalization bounds, loss functions
• (Interim!) CEO of NICTA
– Connect the different parts of ICT research
• Leader of ACOLA project
on Technology and Australia’s future
– Understanding technological change, its mode, its
impact, and its prediction
• This talk: synthesize these diverse perspectives to
(hopefully) provide a stimulation for the future
• Analogies and questions; not many answers…
My Contextual Tripod
3. • Technologies help us do stuff
• They are a means to an end
• Different technologies have things in common
• Predictive technologies can learn from older
technologies
ovement. The illustration shows the size of the machine compared to an operator.
Around 1860, the pages of machine tool books, magazines, and catalogues included
thes, milling machines, drilling machines, grinders, and boring machines. This gives
ome idea of how these types of machines had greatly evolved over a century.
Figure 7.14 illustrates a drilling machine, which is capable of making holes at an
ngle by partially rotating the tool. The hole height could also be changed. Even the
ntire body of the machine could be rotated as a characteristic that was extremely
seful for large machines that were difficult to move. It is evident how machine tools
creased in complexity over the years with the mechanisms that added accuracy,
omfort, and efficiency to the machine operation and application goal.
On Domestic Apparatus
4. Predictive
technologies
• Automation and Control
• Prediction is central to
control
• New control
technologies had
huge impacts on society
• Modern predictive
technologies will too
9. • Technology is not applied science – the
science usually comes afterwards
• “General purpose technologies” steam
power, railways, electricity, IT have driven
industrial revolutions – but they are slow
• Predictive
technologies
next?
How technologies change
10. Infrastructure is slower
• Interdependence slows things down
• What are the interdependencies and
infrastructure for predictive technologies?
11. The evolution of
technology
• Not goal directed
• Source of variation
• Recombination
• Source of selection
• Unpredictable
• Best model of
technological change
13. Technological Inertia, momentum,
path-dependence and lock-in
• Inertia caused by
many things
– Skills
– vested interests
– Interdependence
• Path-dependence and
lock-in
– QWERTY
• Standards can be
good or bad!
14. Prediction of technology
One can accurately predict aspects of technology
Nagy B, D.F.J., Bui Q. M., Trancik J. E., Statistical Basis for
Predicting Technological Progress. PloS one, 2012.
15. But one can not predict impact
• Evidence is from past predictions
of a future which has now arrived
• When you look at all predictions
(and not cherry-pick), everyone
does terribly
16. George Wise, The Accuracy of Technological Forecasts, 1890-
1940, Futures, 411-419, October 1976.
22. Systematization of machines
177On the Systematisation of Machine Study
Fig. 8.8 Classification of movements “Essay on Machine Composition” by A. Betancourt and
J.M. Lanz [73]
23. Cataloging and componentisation
dition, many authors successfully strove to update the theory of machines
hanisms, by considering the acquired technological innovations and experi-
new analytical and design procedures. Examples of this fecundus activity
works of Bourguignon (1906), Franke and Oldenbourg (1930), Bricard
and Grubler (1917), to name but a few.
Mechanisms (From “Kinematische Modelle nach Prof Reuleaux”, Voight (1907) [136])
8 A Vision on Machines
8 “Machine Kinematics and Dynamics” by A. de Lamadrid and A. de Corral (1948) [70]:
t cover; (b) a page
31. Particular
challenges for
predictive
technologies
• Dependence on training data
– Induces a complex dependency that is hard to manage
• How to do experiments
– Tied up with figuring out what one is trying to achieve
• Problem orientation (or lack thereof)
– no different to other technologies
– it is what technologies do, rather than how they do it
that (most) people care about
35. Systems
• Unlike Tony
Stark, no one can
build an entire
system
• We collectively
build them
• The whole point
of APIs!?
36. Modularising predictive
technologies
The only essential knowledge pertains to the
relatedness of things – Jorge Luis Borges
What can be studied is always a relationship or
an infinite regress of relationships. Never a
‘thing’ – Gregory Bateson
• Machine Learning theory
– Functional analysis (functions as points)
– Problems as points
– Relations between problems
– Focus of my technical work
• Predictive APIs
– What to standardise
– How to do it
– Dangers of getting it wrong
37. Data analytics as a service?
• Data big and small everywhere
• Can not necessarily bring the data to one place
• Output of analytics is more data
• Ease of mash-ups
• Allow of plurality of techniques
• State is the enemy – want statelessness
• REST as a solution?
38. REST – Representational State Transfer
An architectural style for data exchange. It is
– Client-Server
– Stateless
– Cacheable
– Layered
RP1 The key abstraction of information is a resource, named by
an URL.
RP2 The representation of a resource is a sequence of bytes, plus
representation metadata to describe those bytes.
RP3 All interactions are context-free.
RP4 Only a few primitive operations are available.
RP5 Idempotent operations and representation metadata are
encouraged in support of caching.
RP6 The presence of intermediaries is promoted.
39. Computational REST (CREST)
An architectural style for computational exchange.
Principles:
CP1 The key abstraction of computation is a resource, named by an URL.
CP2 The representation of a resource is a program, a closure, a continuation, or
a binding environment plus metadata to describe the program, closure,
continuation, or binding environment.
CP3 All computations are context-free.
CP4 Only a few primitive operations are always available, but additional per-
resource operations are also encouraged.
CP5 The presence of intermediaries is promoted.
CREST directly enables computational migration and mobile code.
Facilitates mashups, derived-mashups, and higher order mashups
Superseded by COAST
40. PSI – a start of ML as a service – see
James Montgomery’s talk
41. Other challenges for
predictive technologies
• Privacy, security and all that
• PCAST: information
accountability
• UX for analytics and uncertainty
• Skills
• Deploying predictive
technologies in all sectors of the
economy
Crashlytics
Nexage
Yozio
Kochava
Comscore
Nativex
Mobileaptracker
Google Ads
Appsflyer
Mopub
mDotm
GreyStripe
Millennial Media
Tapjoy
Hockeyapp
Trialpay
ThreatMetrix
Chartboost
Flurry
Google Analytics
InMobi
Adjust
Crittercism
Vungle
Bugsense
Jibro
Pintrest Pet Rescue
Saga
SongPop
MapQuest
Despicable
Me
Avast Mobile
Security
Draw Something
Free
Candy Crush
Saga
Subway
Surf
Tango
Gmail
Trackers Apps Personal Information
Location
Installed Apps
Android ID
Calendar
Figure 5: An example user who is only having 11 apps ye
connected to 26 trackers
7. CONCLUSION
In this paper we presented a measurement study of track
ing in paid apps by analysing top-100 paid apps from fou
di↵erent countries representing four geographical region
We showed that despite having a di↵erent business mode
paid apps also collect significant amounts of personal info
mation, and can lead to the same level of privacy leakage a
when using free apps. We found that 20% of the paid app
had more than three embedded trackers. Also we showe
that 17 out of top-22 trackers collected some form of persis
tent unique identifiers that allows them to track users acros
apps.
By analysing apps installed by over 300 smartphone user
we showed that 50% of the smartphone users are connecte
to more than 25 trackers. The results indicate that it
important to see the overall personal information flow by a
Project Background
NICTA undertook a pilot with the New
South Wales Environment Protection
Authority (NSW EPA) using data
collected from existing environmental
sensors in the Hunter Valley area.
There are 14 environmental sensors
installed by the NSW EPA that monitor
pollution and atmospheric conditions
throughout the Upper Hunter Valley
area. Data from these sensors is
automatically uploaded to the NICT A
developed EPA Air Quality Prediction
Service. Through the use of machine
learning, the predictive algorithms
are then able to infer pollution levels
up to 24 hours in advance including
a confidence reading on the model’s
current performance.
For coal mining companies an
indication of future air quality means
they could take preventive measures
to reduce pollution, which in the long-
term will maximise production.
What isthe technology?
Based on machine lear ning
sensibilities, the pr edictive modelling
tools of the EPA Air Quality Prediction
Service will ultimately allow
Government to access predictions
of pollution for the next 24 hours via
their handheld or tablet in the field.
NICTA’sunique approach
Sensors and the technology behind
them are already well developed in
Australia and internationally. Many
government environmental agencies
are able to collect data about pollution
levels but are unable to accurately
predict into the future.
The EPA Air Quality Prediction Ser vice
uses advanced analytical techniques
to make the most of the data that is
already being collected.
Collaborators
The EPA Air Quality Prediction Ser vice
modelling software is currently being
piloted with the NSW EPA as an
application delivered via the browser
and available via desktop and mobile
devices.
42. Second order cybernetics & infrastructure
• “Infrastructure is both relational and ecological – it means different things
to different groups… It is also frequently mundane to the point of
boredom.”
• Attributes of infrastructure:
• embeddedness,
• transparency,
• reach or scope,
• learned as part of membership,
• links with conventions of practice,
• embodiment of standards
• built on installed base,
• becomes visible upon breakdown,
• is fixed in modular increments, not all at once or globally.
• Because infrastructure is big, layered, and complex, and because it means
different things locally, it is never changed from above. Changes take time
and negotiation, and adjustment with other aspects of the systems are
involved.
• Nobody is really in charge of infrastructure.
What infrastructure
is needed for
predictive
technologies?
43.
44. • Predictive technologies
are technologies
• Look for lessons from
past technologies
• Technologies for data are
likely the next GPT
– Will underpin the next
industrial revolution
Conclusions
45. • Standards for good ….
– Interoperability
• … and for bad
– Lock-in and standards
wars
• Gateway technologies?
– What would they be for
predictive technologies?
Review of Network Economics Vol.3, Issue 1 – March 2004
The Economics of Standards Wars
VICTOR STANGO*
Federal Reserve Bank of Chicago
Abstract
Policymakers face an increasing number of questions regarding whether markets efficiently choose
technological standards. In this essay I survey the economic literature regarding standards, focusing
on arguments that markets move between standards either too slowly or too swiftly.
1 Introduction
The rapid pace of technological change in the last two decades has highlighted the strong
46. • Lightweight minimal standards –
common economic protocol
– Avoidance of lock-in / winner-take-all
– Building an ecosystem rather than
dominating it
• Other challenges
– ML as a service
– Impacts (differential and otherwise)
• Impossible to predict; need to monitor &
adapt
• Fear of “autonomous technology”
– UX for predictions –tricky!
– What infrastructure is needed for
predictive technologies?