Luís Moreira-Matias. Data Science NOW! (and other not so deep thoughts about AI)

Data Science NOW!
(and other not so deep thoughts about AI...)
Luis Moreira-Matias
Luis.Moreira.Matias@gmail.com
Hamburg, Germany .::. March, 2019
Data Science NOW!
(and other not so deep thoughts about AI)
Luis Moreira-Matias
Ph.d. in Machine Learning
Lead Data Scientist
Keynote Speaker
Head of Data Science @ Kreditech
Note: the statements given during this presentation do not represent the institutional positions of the companies that I work(ed) for.

2 © NEC Corporation 2015Real-Time Mobility Data Mining Luis Moreira-Matias2016Data Science NOW!
Computers do not solve everything...
“The computer is not an
intelligent machine that
helps stupid people.
It is a stupid machine that
only works in the hands
of smart people.”
Umberto Eco
(1932-2016)
Italian Philosopher

Artificial Intelligence = Human Intelligence?
“Humankind has given
itself the scientific name
homo sapiens--man the
wise--because our mental
capacities are so
important to our everyday
lives and our sense of
self.
The field of artificial
intelligence, or AI,
attempts to understand
intelligent entities.”
Stuart Russel
(1962-???)
English Computer Scientist

Artificial Intelligence = Machine Learning?
Planning and Control
Perception MotionKnow. Representation Reasoning
Natural Language Proc. Learning

What is Data Science about?
▌Data Science denotes the process of transforming raw data into
knowledge in an (almost...) automated fashion.
▌Related Disciplines:
Statistical Learning (incl. Bayesian Approaches)
Data Mining and Signal Processing
Mathematical Optimization
...among other Computer Science disciplines
Input
Data Viz. and Exploration
Feature
Construction
Missing Value
Procedure
Feature
Rescaling
Sample
Balancing
Representation
Learning/Feature
engineering
Feature
Selection
Outlier
Procedure
Classification
Post Processing
Procedure
Output
Typical Predictive Data Analytics Pipeline

Where Does Knowledge Comes From?
Tribute to Prof. Pedro Domingos (U. Washington)
Evolution Experience
Culture Computers

Computers are everywhere...so is DS!

What is NOT DS about?
▌“I work with data...so I am a Data Scientist!”
▌Business Planning leveraging on Digital Data = Data Strategy
▌Aggregate and Manually analyze (small) Data for Reporting and Decision
Support purposes = Business Intelligence
▌Processes of extracting, querying and displaying (possible large amounts
of) data = Data Engineering
▌Science of Data vs. Science with Data
“Data Science: (not) the preferred nomenclature” –
Peter Flach (U. Bristol),
Editor-in-Chief of Machine Learning Journal
August, 2017

The III Pilars of a Data-Driven Company in 2019
Data Strategy
Data Science
Data
Engineering

Why cant we just mix everything?

Why cant we just mix everything?
Job Specs of a Lead Data Scientist (based on a real world case)…
REQUIREMENTS
▌ Capable of leadership (influence management) and pragmatism;
▌ 5+ years’ experience prototyping classification and regression models using machine software such as scikit-learn, R, MATLAB,
Octave, Weka, Mahout, etc;
▌ Experience with large-scale log processing or big data including but not limited to Elastic MapReduce, Hadoop Streaming,
Pig, Hive, Spark, etc;
▌ Fluency in a range of scripting languages, operating systems, and software platforms including but not limited to Linux, Redis, Java, MySQL,
Python, Pandas, AWS;
▌ Experience in Data Architecture, Data modeling, and Database design;
▌ Experience working with relational and non-relational databases to retrieve structured, unstructured, and semi-structured data sets;
▌ Knowledge of business domains (Travel and Hospitality, Finance, Retail, etc.);
▌ Track record (4+ projects) of working directly with the customer (i.e. regularly facing customer directly) both remotely and on-site;
▌ Presale activities with exposure to customers on billable accounts;
▌ Demonstrated experience with full life cycle of software development processes like RUP or "Waterfall", Agile (SCRUM, XP, etc.).
Experience advocating and establishing an appropriate implementation methodology to successfully develop and deploy the
solution;
▌ Demonstrated experience in solution cost estimation (including tools, tasks, complexity, labor and time) at coarse grain and fine grain levels,
with supporting material evidence;
▌ Demonstrated experience in validating the overall solution from the perspective of performance, scalability, security and capacity.
▌Ability to FLY

So...why are companies doing it?
▌Scarcity of Resources
Adoption still low: only for 12% of use cases defined by 3000 companies [1]
More than 40% of organizations using DS say “the lack of adequate skills” is a challenge [2]
Expected market growth 2017-2022 @ 40% CAGR to $8.8B for ML alone [3]
“Citizen data scientists” market will grow 5x faster than “Professional data scientists” [4]
▌If everyone does it...why cant we too? Just add up “.deep” or “.AI”...
ML-as-a-service: AWS’s machine learning, IBM’s Blue Mix, BigML, Theano, ...
Sales points: Deep Learning sounds sexy, looks sexy and requires tons of resources...justifying sales
figures that would look impossible otherwise - even if they are impossible anyway;
▌Reality...
“We have lots of data but we are not able to use ML
or hire experts to do it so.”
(Rome Transit Agency, Italy, April 2017)
[1] AI The Next Digital Frontier?, McKinsey Global Institute, July 2017
[2] How to do Machine Learning Without Hiring Data Scientists, Gartner, February 2017
[3] Machine Learning Market by Vertical, M&M, September 2017
[4] Predicts 2017: Analytics Strategy and Technology, Gartner, November 2016
[5] Data Scientists Automated and Unemployed by 2025, Data Science central, July 2017

DS Toolset
Data Science

Wait a sec...was not Deep Learning the ultimate solution?

The AI/Deep Learning hype – Why and How?
▌Early 201x - AI Challenges such as
P e r c e p t i o n a n d K n o w l e d g e
Representation are still in dark age!
▌Then, at NIPS 2012...
▌Data, data, data....
▌CNNs + GPUs = $$$
▌It does not overfit ...
...just go deeper or
...add more data!

What is Deep Learning really about?
▌Deep Learning is a first step to perform End-to-End
Learning...wait a sec, what is End-to-End Learning?

▌Deep Learning is a first step towards End-to-End Learning...wait
a sec, what is End-to-End Learning?
Input
Feature
Construction
Missing Value
Procedure
Feature
Rescaling
Sample
Balancing
Representation
Learning/Feature
engineering
Feature
Selection
Outlier
Procedure
Classification
Post Processing
Procedure
Output

▌Deep Learning is a first step towards End-to-End Learning;
▌Feature Engineering (Representation) + Classifier (Decision
Boundary) in a single optimization function!
▌Proeminent Examples:
CNNs (SoA performance in Vision)
LSTM (arguably SoA performance in NLP).
Input
Feature
Construction
Missing Value
Procedure
Feature
Rescaling
Sample
Balancing
Representation
Learning/Feature
engineering
Feature
Selection
Outlier
Procedure
Classification
Post Processing
Procedure
Output

What is Deep Learning really about? – Example with CNN
REFERENCES:
• CS231n Convolutional Neural Networks
for Visual Recognition, Stanford
• Clarifai / Technology
• Deep Learning Methods for Vision,
CVPR 2012 Tutorial

Pitfalls of Connectionist Approaches to Deep Learning
▌Data, data, data, data....and a lot of human labour to label it;
▌Long training times on expensive GPUs;
▌The modelling-to-production time is large and expensive!
▌Lacks a theoretical guarantee of convergence;

Don’t they know that? Yes, but ... 
Ref: David Perkins, The Economist “The world’s most valuable resource is no longer oil...” Accessed March 2018

A note on alternative Deep Learning architectures
▌Zhi-Hua Zhou recently proposed gcForests, a deep learning approach based
on decision trees;
▌It operates by combining random forests with fully randomized trees;
▌It requires less data, less hardware and can be easily fully parallelized;
▌It has lower sensitiveness to hyperparameter settings;
▌It does not overcome CNNs/LSTMs...but it presents a top/competitive
performance in some image/text datasets for classification;
Ref: Zhi-Hua Zhou et al - Deep Forest: Towards An Alternative to Deep Neural Networks (2017, arxiv)

Fiction vs. Reality
▌The right ML algorithm still depends on the task at hand;
▌More than 50% of winning solutions of Kaggle Competitions in the last 5
years did not included any neural network;
▌There is no free lunch – deep or not 
Ref: What algorithms are most successful in Kaggle? (Accessed 03/2018)

Looking ahead: Wait a sec...why can’t a computer do that?
“The computer is not an
intelligent machine that
helps stupid people.
It is a stupid machine that
only works in the hands
of smart people.”
Umberto Eco
(1932-2016)
Italian PhilosopherWhat if we can
CHANGE that?

Data Preprocessing
Feature Selection;
ML algorithm selection/model family;
Hyperparameter Tuning;
Post processing
Model Evaluation and Interpretability;
What is AutoML?
▌ Machine Learning(ML) have been applied to solve relevant problems in multiple
application areas. Successful examples can be listed as:
Credit Risk Modelling (Banking and Fintech);
Personalized Advertising (Retail and Media);
Resource Allocation and Planning (Transportation and Energy);
Client Churning Prediction (Telecommunications);
▌ However, this success is largely based on human ML experts to perform multiple tasks:
AutoML aims to remove the
human-in-the-AI-loop
by progressively automatizing ML applications.
Research Disciplines:
• Regression;
• Bayesian Optimization;
• Meta-Learning;
• Transfer Learning;
• Combinatorial Optimization;

AutoML Success...hum, notable Cases

Banked Customer Underserved Customer
Existing Credit History
Confirmed Address and Identity
Little or no Credit History
No permanent address
High- to middle class income
Regular, fixed-term employment
Regular income pattern
Low- to middle class income
Self-, Freelance or seasonal
employment
Irregular income pattern
Traditional credit underwriting is broken

Predictive & Prescriptive Analytics at Kreditech
– Automated ML-
based on-line
marketing: SEA,
Affiliate, RTB,
Retargeting
– Selective offline
campaigns (TV)
Partnerships with
local players
– Direct integrations
with Partners (POS
as funding source)
– End-to-end fully
automated
scoring decision
engine based on
machine learning
– Includes
Individual
components for
Credit Scoring,
Pricing and
Affordability;
– Credit/Offer
decision up to 60
seconds
– Pay-out to bank
accounts, POS
merchants
– Average time from
approval to payout
<15mins
– Centrally serviced
in-house from call-
center in Romania
– 24/7 availability
– ML-based traffic
estimation and
response time
minimization
– Semi-automated
ML-based
collection engine
recommends the
best collection
strategy (series
of actions)
– Pre-collection
managed by in-
country collection
teams
– Final collection
ousourced (debt
sales)
– Digital account
provides added
benefit and
supports customer
retention
– Automated ML-
based CRM
engine identifies
and promotes
cross- and
upselling
opportunities and
maximizes
retention rate
SCORING /
UNDERWRITING
PAYOUT
CUSTOMER
SERVICE
CROSS-
SELLING /
RETENTION
COLLECTION
CUSTOMER
ACQUISITION
APPLICATION
– Fast and
convenient
application
process
– 100% online /
mobile, paperless
– Fully automated
processing
– Conversion–Risk
trade-off
optimization
Optimized Automated End-to-End Process Flow
$

More than 4M individuals scored and more than €500M originated (2018)
Spain:
1.3m applications scored
€158 millions originated
Poland:
1.7m applications scored
€230 millions originated
India:
6k applications scored
€700 thousand originated
Russia:
600k applications scored
€100 million originated
Long-term loans
Shot-term loans

Data Science team @ Kreditech - Who are we and what we do?
Abhishek Verma Farouk Sellami
Fuzhen Wang Julien Henry
Luis Moreira-Matias Morteza Alamgir
Panagiotis Katsas
Heschmat Seyedi
Ragunath Sivakumaran
Yiannis Gatsoulis
Current R&D Topics:
- AutoML for Sampling, Feature Selection
and Hyperparameter Tuning;
- Affordability w/ ML-based Survival Analysis;
- ML Model Uncertainty Estimation;
- Reject Inference w/ Semi-Supervised
Learning and Evolutionary Optimization;
- Fraud Detection w/ Supervised Learning;
We are
Hiring!!!

34 © NEC Corporation 2015Real-Time Mobility Data Mining Luis Moreira-Matias2016Data Science NOW! Luis Moreira-Matias
1. Nowadays, Artificial Intelligence is still far from Human Intelligence;
2. DS is about to transform raw data into knowledge automatically;
3..Data Science is one of the most thrilling industries worldwide.
Why? Computing Power and loads of digital data available!
4. The lack of experts pushes organizations to stretch its definition way beyond
its real scope;
5. Deep Learning? Don’t fall by jumping in the hype.
Neural nets are super cool...but they are just yet another tool.
6. Be aware: Automation is the new battleground!!!
7. DS @ Kreditech: 6+ yrs. and 600M of 100% human-free credit underwriting globally!
And we have done this without using a single Deep Learning model in production... :o
▌7 Tomatoes or ... Claps Time!

Luís Moreira-Matias. Data Science NOW! (and other not so deep thoughts about AI)

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à Luís Moreira-Matias. Data Science NOW! (and other not so deep thoughts about AI)

Similaire à Luís Moreira-Matias. Data Science NOW! (and other not so deep thoughts about AI) (20)

Plus de IT Arena

Plus de IT Arena (20)

Dernier

Dernier (20)

Luís Moreira-Matias. Data Science NOW! (and other not so deep thoughts about AI)