Machine learning in general and deep learning in particular are driving major advances for a wide range of specific finance use cases. This talk will outline how enterprise-wide learning loops will extend these point success to a coherent AI strategy and also show what other elements are required for success, using real-world examples at Prudential plc.
Watch the presentation here: http://videos.re-work.co/events/36-deep-learning-in-finance-summit-london-2018
Horses for Courses: Deep Learning Beyond Niche Applications
1. Michael Natusch, Global Head of AI, Prudential plc
Deep Learning in Finance Summit, 15 March 2018
Horses for courses:
deep learning beyond niche applications
2. 2
Prudential – providing financial security since 1848
2
Asset
Management
Asia & Africa US UK & Europe
Insurance & Savings RetailInstitutional
Distribution &
Asset Allocation
how can we productionalise deep learning here?
3. 3
How can we build an AI product?
intelligent
agent
+
frictionless UX
data UI
+
learning loop
Social
Agents
Customers
Partners
automated
Actions
4. 4 2010 2020
deep
learning
probabilistic graphical models
random
forests XGBoost
Source: Wolfgang Ertel, "Introduction to Artificial Intelligence", Springer Verlag, 2011
algorithms
5. 5
Hospital
what algorithm when??
unit cost
($/transaction)
volume
(# transactions)
Google
correlation is good enough
need causation
probabilistic
graphical
models
gradient-boosted
decision trees
deep
learning
6. 6
some of our recent AI-related releases
6
unit cost
($/transaction)
volume
(# transactions)
service
bot
agent
productivity
robo
advisor+£1m AUM/wk
record
retrieval
75%
automation
audio
search
94% automation
@ 90+% accuracy
claims
mgmt
handwriting
recognition
7. 7
Claims management
+
frictionless UX
new claims
+ historic
claims data
+
learning loop
Social
Agents
Customers
Partners
automated
Actions
2 model
ensembles
mobile UI
(incl. chatbot)
l
10. 10
Neural network architecture
Four convolution layers followed by two fully connected layers form the
architecture of the character recognition model
48x48
Grayscale Image
100
200
300
400
1000
5507
Max Pooling
Max Pooling
Max Pooling
Max
Pooling
Drop
out
Soft-
max
Convolution
Layer 1
Convolution
Layer 2
Convolution
Layer 3
Convolution
Layer 4 Fully
Connected
Fully
Connected
11. 11
Field Image Output Correct Characters
Accuracy (1=Correct,
0=False)
Total / Correct Correct Percentage
Name of
Policyowner
1 1 1 3/3 100.00%
Name of Life
Assured
0 0 1 3/1 33.33%
Name of
Employer
ABC 0 1 1 1 1 1 1 1 1 9/8 88.89%
Residential
Address A
1 1 1 1 1 1 1 1 1 1 1 1 1 1 14/14 100.00%
Signs and
symptoms
1 1 2/2 100.00%
Name of
Physician /
Hospital
1 0 0 1 4/2 50.00%
Name of
Physician
1 1 2/2 100.00%
What is / are the
underlying
cause(s) for final
diagnosis?
1 1 1 0 4/3 75.00%
41/34 82.93%
Chen Dawen
Chen Dawen
ABC Logistics Limited
Hong Kong Happy Garden,
11th floor, Room A
stomach ache
Mary Hospital
Li Wen
food irritation
Performance on Pru Hong Kong test form
12. 12
Chinese English Digits
400GB of labelled handwriting data
covering the 5,507 most frequent
characters
Microsoft/NIST MNIST
Python - Tensorflow Python – Tensorflow/Microsoft API Python - Tensorflow
CNN CNN + Microsoft API CNN
Prudential Hong Kong Hospital
Claim Form
Prudential Financial Planning Form Prudential Financial Planning Form
85% 89% 98%
Training Data
Framework
Model
Architecture
Test Data
Accuracy
Model summary
14. 14
Prototypes
a Alpha prototypes are typically created in a hothouse
to quickly prove user experience and functionality fit
a+ Alpha+ prototypes are matured to mimic real systems
and responses, and are ready to be piloted with
colleagues for initial feedback
b Beta prototypes are one step away from production,
having been rapidly iterated upon with feedback from
customers and colleagues
W Omega projects are live, operational, robust and
defect free
15. 15
Prototyping accuracy
Case Load Min AUC equivalent
a ~ 102 ≥ 50%
a+ ~ 103 68.3% 1 s
b ~ 104 88.8%
Wmin
~ 105 95.4% 2 s
Wmax
~ 108 99.99994% 5 s
16. 16
Potential health insurance use cases for this
• Claims & underwriting
– Automation of existing from processing
– Leading to end-to-end process automation
• Fraud, waste and abuse
– Identify claims patterns across customers, agents, providers and clinicians, leading to
identification of fraud and collusion, wasteful practices and abuses
– Can be extended to a ‘real-time’ system, feeding into claims & underwriting
– Enables a learning loop for the automated process
• Customer insight
– Faster feedback on emerging trends across customers and conditions
17. 17
How can we build an AI product?
intelligent
agent
+
frictionless UX
data UI
+
learning loop
Social
Agents
Customers
Partners
automated
Actions
18. 18
Best practice for implementing AI successfully
• data
• tooling
• infrastructure
• people
• APIs
“technical”
• start small and build stuff
• experimentation is cheap
• collaboration, co-location, cross-
functional, time-boxed
• iterate frequently
• drive actions & learn continuously
from their outcomes
“cultural”
build iteratively improving prototypes and measure their accuracy
19. 19
5 maturity levels for building AI products
1. historic data analytics (including advanced analytics techniques such
as clustering, social media analysis, basic NLP etc.)
2. stand-alone machine learning models that are integrated into the
systems stack and drive actions on real-time data streams
3. predictive models that are exposed via APIs and conversational
interfaces
4. automated, continuous learning loops based on batch model re-runs
and real-time data streams on a personalized user basis
5. learning loops that automatically and continually redefine and
implement product, service and channel offerings – autonomously