Some weeks ago, our ML6 agent Karel Dumon gave a talk at a Nexxworks Bootcamp. During this week-long event, several speakers are invited to take the floor to inspire a heterogenous group of (senior) business people from a wide range of industries. On the third day, Artificial Intelligence was planned. A broad intro to AI and ML was given by prof. dr. Eric Mannens, after which Karel provided the audience with some hands-on insights through use cases.
3. Nexxworks Bootcamp Ghent - 27/09/2017
Team of Data Engineers, Data Scientist & Machine
Learning Engineers
Closing the gap between lots of data - lacking insights
Robust & agile ML solutions through scalable APIs
Premium Partner of Google Cloud
8. Confidential & ProprietaryGoogle Cloud Platform 8
2012 20132002 2004 2006 2008 2010
GFS
MapReduce
Bigtable Colossus
Dremel Flume
Megastore
Spanner
Millwheel
Pub/Sub
F1
2016
Dataflow
TensorFlow
Google’s 15+ years innovation in data
9. Confidential & ProprietaryGoogle Cloud Platform 9
Decrease the innovation gap
2012 20132002 2004 2006 2008 2010
GCS
Dataproc
Bigtable GCS
BigQuery Dataflow
Datastore
Spanner
Dataflow
Pub/Sub
F1
2016
Dataflow
Cloud ML
NoSQL
11. 11
Rapidly Accelerating Use of Deep Learning at Google
Number of projects using some form of deep learning
2012 2013 2014 2015
1500
1000
500
0
Used across products:
12. Nexxworks Bootcamp Ghent - 27/09/2017
What’s going on?
The cloud is very good at
handling, storing and manipulating
large volumes of data.
What we really care about now is
understanding the data.
17. Nexxworks Bootcamp Ghent - 27/09/2017 17
The old, algorithmic approach
“apple”
“orange”
“banana”
IF (round) THEN
IF (orange AND coarse) THEN
“orange”
ELSE IF (green AND smooth) THEN
“apple”
ELSE IF ...
...
ELSE IF …
“banana”
18. Nexxworks Bootcamp Ghent - 27/09/2017 18
Let the machine find the rules
“apple”
“orange”
“banana”
?
22. + ‘The Next Big Thing’
+ Sentient AI in the next 10 years
(‘The Singularity’)
+ Will put humans out of a job
+ Foolproof
MACHINE LEARNING
in popular culture
23. + Been around for 60 years now
+ ‘Sentient next year’, every year,
for the last 60 years
+ AI winters: 1970, 1990, … ?
+ Not foolproof
MACHINE LEARNING
reality
24. 24
A person on a beach
flying a kite.
A person skiing down a
snow covered slope.
A group of giraffe standing
next to each other.
25. 25
A woman riding a horse
on a dirt road.
An airplane is parked on the
tarmac at an airport.
A group of people
standing on top of a
beach.
33. Nexxworks Bootcamp Ghent - 27/09/2017 33
Logistics optimisation
Predicting the time packages stay in customs
+ Better predict airplane cargo loads
for an international courier
+ For each parcel, predict the time
under inspection by customs
(0 min = not picked up)
+ Better prediction → Better planning
Better planning → Fuller cargos
Avoid sudden overload
34. Nexxworks Bootcamp Ghent - 27/09/2017 34
Training the deep learning network (simplified)
A parcel’s
attributes
(weight, size, text,
colours, origin ...)
The predicted
package delay
in customs
>100k examples
Full case: 3 days!
35. 35
The goal was to predict congestion in
London 40 minutes ahead by looking
at the stream of data from induction
loop road sensors.
Congestions usually ripple through the
city, so should be predictable.
Warnings ahead of time could help
prevent gridlocks.
41. Nexxworks Bootcamp Ghent - 27/09/2017 41
Predictive Maintenance: towards a highly dynamic Digital Twin
Design Connect Predict
theoretical model collect data create Digital Twin
Testing generates theoretical
life-time model
- in-house testing
- remaining life-time in
‘ideal conditions’
Machine Learning model
predicts expected failure time
Client receives personalized
model by including custom
environmental factors
- usage, temperature,
humidity…
Cloud computing/storage
STATIC MODEL HIGHLY DYNAMIC ML MODEL
connected device
Digital Twin
ML
43. 43
What did we do?
+ Shipped state-of-the-art system in under one month
turnaround time
+ System is not an ‘Artificial Intelligence’
+ System outperforms most solutions currently marketed as AI
Why use Google Cloud Platform?
+ Access to massive computational resources to train our models
+ Able to dynamically adjust to changing number of requests
+ Leverage Google powered API’s
The solution
62. 62
Entity Extraction
● “I got charged 7 dollars for a Christmas
Card that I never ordered. Plz remove this
from my invoice, my telephone number is
472867486.”
● “My internet connection is no longer working. I
have ADSL smart-50 and my address is NY
city 52 A on 21st.”
Entity Extraction Postprocessing
7 dollars $7.00
Christmas Card Christmas-card
472867486 +32472867486
ADSL smart-50 ADSL smart 50
NY city 52 A on 21st
52A, 21st avenue, New
York
63. Nexxworks Bootcamp Ghent - 27/09/2017
Automatic maintenance
detecting faulty solar panels with drones
64. Nexxworks Bootcamp Ghent - 27/09/2017 64
+ Improve maintenance efficiency of a solar panel company
+ Using drones with thermal image cameras
+ Classify broken panels vs defect free panels
+ Deep convolutional network
65. Nexxworks Bootcamp Ghent - 27/09/2017
Cloud Storage folder
to_classify
Cloud Storage folders
ML6 API
defect
normal
Machine Learning
model on Cloud ML
Dashboard
Upload captured
images to Storage
Results and interpretation on an
interactive dashboard
66. Nexxworks Bootcamp Ghent - 27/09/2017 66
Other ideas?
smart cities? (garbage detection, road conditions, mobile camera’s for anomaly detection...)
69. MEDICAL IMAGE CLASSIFICATION
TECHNICAL DETAILS - DATA
Camelyon16: ISBI challenge on cancer metastasis detection in lymph node
Training & Evaluation: 110 tumor slides, 160 normal slides, 130 evaluation slides
Task 1: whole-slide level prediction binary classification problem
Task 2: find metastasis location segmentation problem
71. MEDICAL IMAGE CLASSIFICATION
TECHNICAL DETAILS - CHALLENGE BREAKDOWN
Not just another standard image classifier…
Network architecture
Training set construction
Computing environment
Post-processing (classification and segmentation)
72. MEDICAL IMAGE CLASSIFICATION
TECHNICAL DETAILS - TRAINING
Patches are used to train a deep learning model
▹ started with simple CIFAR10-model
▹ ended up retraining Inception-v3 to experiment with synchronous updates across
multiple GPUs
Metastasis patches
Normal patches
Deep Learning model
e.g. Inception-v3
73. MEDICAL IMAGE CLASSIFICATION
TECHNICAL DETAILS - MULTIPLE GPUs
Multiple GPUs were used to calculate model
updates (based on TensorFlow tutorial)
▹ individual model replica on each GPU
▹ all GPUs process a batch of data and send
update of model parameters to CPU, which
performs a synchronized update
Approach was extended to Cloud ML Engine GPUs
during an external ackathon to build a dermatology
POC in 1 day (by starting from a.o. pretrained VGG16)
89. Nexxworks Bootcamp Ghent - 27/09/2017 89
Use reinforcement learning to optimize energy usage
Reinforcement learning allows for automatic parameter optimization:
- e.g. energy optimization: let AI-agent control settings to optimize
energy consumption in data center cooling
- See also Google Deepmind
93. Nexxworks Bootcamp Ghent - 27/09/2017 93
Advantages that AlphaGO can leverage
1. Fully deterministic: no noise in the game
2. Fully observed: each player has complete information and there are
no hidden variables. (unlike Poker for example)
3. Discrete action space
4. Each game is relatively short (approximately 200 actions)
5. Target function is clear (win/lose) & fast to evaluate
6. Huge datasets of human gameplay are available to bootstrap the
learning, so AlphaGo doesn’t have to start from scratch
95. Nexxworks Bootcamp Ghent - 27/09/2017 95
“If a top 1% CEO today
understands how software applications get built and
how that changes the way she/he manages,
a top 1% CEO in the near future
will understand how models get built and
how that changes the way she/he builds her/his organization.”
James Cham, Bloomberg Data
100. Tensorflow Training
THANKS!
What my friends think I do
What other computer
scientists think I do
What society thinks I do
What mathematicians think I do What I think I do What I actually do
Talk about AI → ML is a technique that is currently most used in the context of AI
Frame Machine Learning: what is it, why is it working now and not before?
What can it do? Where does it fail?
Cases to get a very good grasp of the concepts; mostly supervised learning
What’s Next: Reinforcement Learning
Close with some tips on how to do ‘AI’ yourself
Note: limit my talk to ML (and not general AI), because that’s our scope atm
Also: not unsupervised learning
Note: limit my talk to ML (and not general AI), because that’s our scope atm
Also: not unsupervised learning
Note: limit my talk to ML (and not general AI), because that’s our scope atm
Also: not unsupervised learning
Note: limit my talk to ML (and not general AI), because that’s our scope atm
Also: not unsupervised learning
When you think of Google, you probably think of this: a simple and elegant search box that delivers an instant answer
But when I think of Google, I think of this: infrastructure, network, datacenters, servers, all designed to enable developers to build innovative software and services
And the reason they had to build all this is because back when they started, the technology they needed simply didn’t exist.
And of course, having built that infrastructure, it also enabled them to build new products like GMail, Maps, YouTube, Chrome
Over the years there have been running some interesting technologies inside of Google.
And they have published papers describing these technologies to the outside world.
Historically, the open source community would read these papers and built out technologies based on the described concepts.
One important thing to note here is that Google has had Hadoop in production since early 2000s. Most enterprises did not start using Hadoop at-scale until at least 2010. That’s 10 years. In 10 years the Dow Jones has lost 50% of companies from its index. What kind of impact does not having a potentially competitive technology for 10 years have on your business?
Well all this has changed with the growth of Google Cloud Platform.
Everything that Google will build for the future will be built with external customers in mind. So this 5 to 10 year gap from when Google has something to when customers have something, is now erased down to 0 - if you are on Google Cloud.
2002: cool stuff but industry lagging behind
Now: open source - research/
MapReduce in 2004: paper that was basis for Hadoop
MapReduce is a programming model and an associated implementation for processing and generating big data sets with a parallel, distributed algorithm on a cluster.
BigTable → H Base
Now: TensorFlow & Kubernetes directly
Google’s case also nicely shows where we are now: understanding the data
That is why wachine learning is everywhere already…
What are some of the ways that Machine Learning is having a significant impact at Google?
All of these examples implemented uses of TensorFlow or our predecessor system
With Machine Learning
We have the general purpose Cloud ML platform which runs TensorFlow models at scale
Then the pre-built ML APIs
Like Natural Language for parsing text and extracting entities, sentiment, etc
Speech API takes spoken word and turns it into text
And finally Vision API - you give it images, it gives you back a ton of metadata, like places, objects, faces, etc.
Computer sets the rule, mini brain, black box
Loads of oranges bananas...
Computer sets the rule, mini brain, black box
Loads of oranges bananas...
How we are about to achieve this? by using Machine Intelligence. (and deep neural nets)Making sense of (lots of) existing data to make predictions about new data.And getting smarter from that new data to make ever better predictions.
Insight is very important to leverage it for your business
Is it just ‘A Machine’?
Luckily there are teams like us to guide you through the AI/ML jungle :-)
AI and ML can really empower your business, help you today, tomorrow, and definitely the day after tomorrow
Become ‘AI first’ and make your company future proof
Start small, learn the basics, build a POC and go on from there
How to model this?
Which attributes?
How to present the input vector to the network?
Which metric to predict? Weight by time? Actual delay?
Which implementation? LSTM, convolutional? How many layers?
Just using
Importance of reusing and retraining data
Text, weight (very important), colours, origin
>100000
in 3 days
Demand & Supply forecasting
Demand & Supply forecasting
Demand & Supply forecasting
A customer collects images of solar panels with thermal cameras attached to drones.We were asked to develop a solution able to classify them as broken or defect free.We want the technicians to spend time efficiently, fixing the broken panels instead of looking for problems.
1000 images
We build full solutions. The model is just 1 part.. It also needs to be retrained
BigQuery!
Other use cases: smart cities? (garbage detection, road conditions, mobile camera’s for anomaly detection,...)
A customer collects images of solar panels with thermal cameras attached to drones.We were asked to develop a solution able to classify them as broken or defect free.We want the technicians to spend time efficiently, fixing the broken panels instead of looking for problems.
+ you don’t need to be an expert, but you might still need one!
Decision making
From simulation to real-life?
Meerdere agents
Toon pacman hierna
It’s not only in robotics!
In our fund, where we invested exclusively in the future of work from our first check, we see a new kind of company leader emerging. She does not see her role in just the old way — deciding and translating decisions into software and people workflows, or even as nurturing her people to do that. She sees her role in a new way — creating the conditions (in talent and data) to build the right models that can make better decisions than any of us, and — critically — understanding when not to trust models and when to update them. This creates an organization that can make decisions more rapidly (and improve more rapidly) than any pointy-haired boss could through tedious training or tenuous anecdotes.
The previous generation understood how to lead people and how to harness IT. The new generation will understand how to manage models, which are the flux capacitor of making software go beyond workflows to decisions.
If a top 1% CEO today understands how software applications get built and how that changes the way she manages, a top 1% CEO in the near future will understand how models get built and how that changes the way she builds her organization.
A whole way of thinking about leadership, and about the economics of companies, with as many maxims and guidelines as the old way, is to come.
We’re waiting for the Peter Drucker of machine intelligence.
James Cham
Partner, Bloomberg Beta