evolving machine intelligence and its influence on risk...
TRANSCRIPT
Evolving machine intelligence and its influence on risk landscapes & analyses
Dr. Jeffrey Bohn, Head, Swiss Re Institute
CDAR Symposium, 19th October 2018
2
Digital Society = People + Data + Networks + Machine IntelligenceWhat risks emerge?
Inspiration from Ludwig Boltzmann (Austrian physicist & philosopher who developed statistical mechanics; coined the term ergodic)
3
In my view all salvation for [machine intelligence] may be expected to come from Darwin’s theory
Boltzmann’s ergodic hypothesis: For large systems of interacting particles in equilibrium, time averages are
close to the ensemble average.
Outline for presentation
• Digital society
• Changing risk landscapes
• Reinsurance
• Evolution of risk modeling
• Machine intelligence
• Machine intelligence’s fragility
• Matching the algorithm/system/process/data to risk-related use cases
• Final remarks
4
Digital society
5
The Digital Society is the result of technology abundance
Semiconductors enabled cheap and abundant computation
The internet enabled cheap and abundant connectivity
Big data is creating cheap and abundant information
IoT is enabling cheap and abundant sensors
…
Machine intelligence is enabling cheap and abundant predictions
In digital societies, if data is the oil, Machine Intelligence is the refinery
6
Rise of the machines
7
Overcrowding AI WinterMachine Learning
Deep Learning Meta-learning
Early 1980’s Early 1990’s Early 2000’s Late 2000’s Now – Future
Primitive AI that required many
sets of rules to perform basic functions
No substantial innovation; caused by weak computer
hardware and lack of data
Exploration into learning systems that are fed large amounts of data
Innovating machine learning systems to work unsupervised
and faster
Current/ongoing development of
learning systems aim to mimic humans
Rules basedPattern matching
Linear modelsClustering
RegressionsSmall ANNs
Non-linear models
Sigmoid/Back prop.Deep RNNs
Boosting
Boosting, Boltzmann machines, GANs
Projective simulationQuantum-inspired
Evolutionary
Perform basic tasks normally done
by humans(CPU Chess Players)
Stalled while awaiting more advanced
computer systems and data
Create self-learning systems that feed
on data (Stock trading algorithms)
Self-learning systems that improve unsupervised
(Self-driving Cars)
Systems that accurately mimic human behavior
(Chatbots)
Note: AI: Artificial intelligence; ANN: Artificial neural network; RNN: Recurrent neural network; GAN: Generative adversarial networks
Description
Models
Objectives
Rise of the networks
Description
Early networks were funded by a few governments such as the U.S.
Migration to dial-up internet and
proliferation of client-server
networks
Internet penetration in
developed world exceeds 50%
Change economics of networked
computing in terms of fixed investment
Further change of networked
computing to further reduce
investment, add IoT
Examples Arpanet Top-level domains e.g., .com take hold
Appliances shipped with IP addresses
AWS, GCP, and Azure
Make appliances “smart” e.g., Tesla
Objectives
Connect researchers & military
Connect universities &
companies
Connect companies,
governments, universities &
consumers
Create more elastic networks that
extends networked computing
Move machine intelligence out to network edges to increase speed &
create new “intelligence”
GovernmentWorldwide Web & Client-servers
Ubiquitous Internet/ IoT
Cloud Edge & Fog
Early 1980’s Early 1990’s Early 2000’s Late 2000’s Now – Future
8
Data is the new oil: Crude product multiplies in value after refinement
Generate & collect
Organise & curate
Analyse & transform
Deliver& multiply value
Platform (eg. Smart City)
IoT 5G Connectivity Big Data Cloud computing Human + Artificial Intelligence Autonomous systemsRobotics
9
Changing risk landscapes
10
• Insufficient investment in data collection & curation systems
• Digital “exhaust” and “hygiene”
• New networks overlayed on legacy networks create gaps and vulnerabilities
• Differing value/regulatory systems underlying society’s digitization create disconnects
• Over-dependence on automated/machine-intelligence-enabled systems leads to blind spots and unnecessary risk
• Vulnerabilities arise from IoT-enabled devices with outdated security features
• Mismatching digital tools with a given use case leads to poor implementations & vulnerabilities
Cyber risk and algorithmic malpractice increasingly becoming most important risk categories
Digital society and new risk landscapes
11
Lessons from technological change in a different industry
12
Source: Gartner Inc. (2015)
Sale of Cameras
13
Changes arising from distributed ledger technology (DLT)
Internal improvements
Enhancement ofinsurance value-chain New value-chains
End-customer driven initiatives
New players creating a DLT-based alternative value-
chain or market for distribution of re/insurance
services
End-customers in non-insurance verticals
organizing DLT cooperatives for aggregation of shared
services including insurance
Organizing members of existing insurance value-
chain into a DLT cooperative for driving data
standardization, aggregated data access and shared
processes
Enhancing internal business process
efficiencies by hosting shared services on internal
DLT platforms
End-customer
Capital Insurer
Broker Reinsurer
End-customer
End-customer
End-customer
Service
Insurance
Service
End-customer
Capital
new player
new player
Incremental efficiency gains
Business model disruption
What is Cyber Risk?
14
Accidental breaches of security / Human error.
Unauthorized and deliberate breach of computer security to access systems.
IT operational issues resulting from poor capacity planning, integration issues, system integrity etc.
How severe is this risk?
15
1.9B records stolen in first 3 quarters of 2017 compared to 600M in entire 2016
500M forms of malware (400k created everyday)
1M identities breached per breach
Probability of breach in a year is close to 100%
The global cyber insurance market is still small…
271
196
97
2.9
Motor
Property
Liability
Cyber
North America
199
41
19
0.2
Motor
Property
Liability
Cyber
Asia Pacific184
117
42
0.4
Motor
Property
Liability
Cyber
EMEA
Source: Swiss Re Institute
Insurance premium in 2017 per LoB, in USD bn
16
…but expected to grow rapidly
17
Expected global cyber insurance premium volume (2015 – 2025), in USD bn
Source: Swiss Re Sigma No 1/2017, Cyber: getting to grips with a complex risk
Allianz
Aon
PWC
Advisen
ABI
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
202520202015
20% CAGR 32% CAGR
44% CAGR
21% CAGR
15% CAGR
37% CAGR
(2015-2020)
(2015-2020)
(2015-2020)
(2015-2020) (2020-2025)
(2015-2020)
Cyber-insurance market is expected to grow more rapidly than other lines of business. Currently, market focus is on the SME segment. However, first cyber insurance solutions for individuals are also entering the market.
Global insurance premium growth (1960 – 2017)
Source: Swiss Re Sigma No 3/2018, World insurance in 2017
1.5%
Reinsurance
18
Reinsurance is a catalyst for economic growth.
19
Reinsurers absorb shocks, support risk prevention and provide capital for the real economy
Activity Benefits
Risk transfer function Diversify risks on a global basis Make insurance more broadly available and less expensive
Capital market function Invest premium income according to expected pay-out
Provide long-term capital to the economy on a continuous basis
Information function and knowledge
Price risks Set incentives for risk adequate behavior
20
Assets and Underwriting steering have similar basic building blocks and should be done jointly
common drivers
Current Portfolio
Forecasts
Management Actions –shaping the target
Target Portfolio
NAV by Asset Class and segment
Expected Returns
Shift across assetclasses; IR Duration changes
Long-term Strategic Asset Allocation (SAA)
Expected premia & claims by UW portfolio segment
Forward-looking Views (FLV) on exposure, loss and premium rate trends
Organic growth, Pruning, Hedging, Initiatives, Large Transactions
Target Liability Portfolio or BU Plan
• Capital allocation: Choose/avoid outperforming/underperforming insurance portfolio segments (IPS)
• Risk selection: Select risks within each IPS
• Strategic asset allocation: Choose/avoid outperforming/underperforming asset classes
Forward-looking modeling is key to improving a reinsurer’s performance
21
22
More examples where machine intelligence changes insurance
Forward-looking modeling of risk pools
Incorporating unstructured data into business and capital steering
Tracking natural catastrophe damage in real time
Assessing damage
Automated underwriting
Improving customer targeting
Parametric insurance contract implementation
Intelligent automation & robotic process automation (RPA) for underwriting and claims processing
Chatbots for customer support
Natural language processing applied to contract review
Evolution of risk modeling
23
“Risk” definitions matter
• Symmetric return/loss distributions with parameterizable distributions based on existing data– “known unknowns”
– Volatility measures may be sufficient
– “Beta” risk should be the focus i.e., allocation more important than individual exposure selection
• Asymmetric return/loss distributions with changing distribution parameters with sparse data– “partially known unknowns”
– Focus shifts to tail risk/expected tail loss
– Diversification opportunities continue even as portfolio becomes quite large
– Active management may be better compensated given huge savings to avoiding tail events
• Emerging risk means data availability may be so sparse as to make any parameter estimation infeasible–“unknown unknowns”
– Ambiguity creates risk that cannot be managed using traditional approaches
– “Structural” model based on subject matter experts can provide some guidance, which cast some light on unknown
24
25
Expected
Loss
Risk Contribution
Probability
0 Exposure Loss
No Loss
Expected Tail Loss
(Extreme Loss)
RISK CONTRIBUTION (RC) & EXPECTED-TAIL-LOSS CONTRIBUTION (TLC)
RISK CONTRIBUTION
• Exposure’s contribution to portfolio’s
volatility
• Relates to shorter-term risk
EXPECTED TAIL-LOSSCONTRIBUTION
• Exposure’s contribution to extreme
loss possibility (i.e. the “tail”)
• Relates to rare, but severe losses
* Volatility Multiple
* Analytical solutions
Simple* Flexible
* ETL contribution
Simulated* Dynamic
* Individual exposure differentiation
Stratified
• Defining extreme-downside, scenario categories:
– Black swans: Unknowable given current information set and virtually impossible to predict
– Gray rhinos: Highly probable and straightforwardly predictable given current information set, but neglected
– Perfect storms: Low probability and not straightforwardly predictable given the outcome results from interaction of infrequent events
• Scenario-based analyses vs. forecasts
• Deeper analyses of underlying assumptions, relationships, and data
• More focus on tools/processes to manage multiple sets of scenarios and analyses across time
• Renewed efforts to enforce preproducibility, reproducibility, and out-of-sample testing
• Process management systems with robust audit logs are more important than ever
Black swans, gray rhinos, and perfect storms
26
• Sources of non-stationarity/ non-ergodicity
– Structural changes in the real economy
1. Digital ecosystems
2. Increased concentration within sectors
3. Information & communications technology
– Structural changes in the financial economy
1. Near-zero interest rates
2. Inflation
3. Globalization of capital markets
• Ergodic systems
– Closed
– Low dimensional
– Not evolving, but can be cyclical
• Non-ergodic systems
– Open
– Generate new information all the time
– Learn and adapt over time (e.g., Darwinian evolution)
– Past data mostly not useful
– Purely inductive models are mostly ineffective
Challenges specific to financial market and insurance risk modeling
It is far better to have absolutely no idea of where one is– and to know– than to believe confidently that one is where one is not.-- Jean-Dominque Cassini, astronomer, 1770
27
Marrying qualitative and quantitative data
• Roughly 90% of available data are qualitative and unstructured e.g., articles, blogs, e-mail, regulatory filings, slide presentations, social media, etc.
• Quantitative data may not reflect all forward-looking risks (e.g., Environment, Social, Governance-- ESG)
• Transforming qualitative data into indicators and combining in some way (e.g., shading, weighted combinations, etc.) with quantitative data may be a path to improving existing models
28
Misusing Occam’s razor
William of Ockham’s actual quote (most likely): Numquam ponenda est pluralitas sine necessitate
[Plurality must never be posited without necessity]
Simplicity in concept may belie complexity in reality– e.g., biological evolution, construct an optimal portfolio, optimally allocate capital to insurance portfolio segments, etc.
29
Machine intelligence
30
• Artificial: Human-made, contrived, not natural, not real
• Intelligence: Learn and apply knowledge or skills, solve problems, ability to reason/ plan/ adapt/ respond/ think abstractly
• Artificial intelligence: Ability to simulate human intelligence– is this all?
Are these definitions adequate? How we define machine intelligence materially influences the tools we create
“Viewed narrowly, there seem to be almost as many definitions of intelligence as there were experts asked to define it.” Sternberg
How we define “artificial” and “intelligence” will influence research and development in machine intelligence
YOU ARE HERE
The future will be Human Intelligence augmented by some kind of Machine Intelligence
31
32
Machine intelligence taxonomy
Artificial intelligence: Mimic human intelligence and possibly go beyond
Artificial general intelligence: Possibly self-aware intelligence
Machine learning: Data-dependent calibration
Deep learning: Model-free, data-dependent calibration
Meta-learning: Learning how to learn
Cognitive computing: Simulate human-brain processes
Augmented intelligence: Human assistants
Expert systems: Advice systems using knowledge databases
Robotic process automation: Roboticized systems that replicate repetitive processes
Intelligent automation: Hybridized processes that use automation to better leverage human productivity
Six major categories for applying machine intelligence in business applications
Machine Intelligence for
Business Applications
Machine Learning Platform (ML)
Deep Learning Platform (DL)
Natural Language Processing (NLP)
Internet of Things (IoT)
Recognition Technology
Decision Management
Supervised Learning: Regression, Classification, Dimension Reduction..
Unsupervised Learning: Association rule learning, clustering…
Reinforcement Learning
Neural Network/ Convolutional Neural Network
Speech Recognition/Question Answering
Natural Language Generation
Virtual Agents
Digital Twin/AI Modeling
Content Creation
Peer-to-Peer Networks/ Edge Computing
Emotion Recognition
Image Recognition
Decision Support System
Infotainment System
33
• Boltzmann machine: Stochastic recurrent neural network that can be seen as the stochastic, generative counterpart of Hopfield nets. Maximizes the log likelihood of given patterns exhibiting Hebbian engrams. Threshold is binary. Finds hidden relationship layers.
• Hopfield networks: Form of recurrent artificial neural network that is content-addressable with binary threshold nodes.• Hebbian engram: Two cells (or systems of cells) that are repeatedly active at the same time tend to become
“associated.”• Dynamic Boltzmann machines (DyBM): Maximize log likelihood of time series with property of spike-timing dependent
plasticity (STDP) [amount of synaptic strength change depends on when two neurons fired.] Threshold is binary.• Gaussian Dynamic Boltzmann machines: Extend DyBM to deal with real values capturing long-term dependencies
(similar to VAR, which is a special case). Can add an RNN layer to induce hidden, high-dimensional non-linear relationship features.
Ackley, David H., Geoffrey E. Hinton, and Terrence J. Sejnowski, 1985, “A learning algorithm for Boltzmann machines,” Cognitive Science, 9(1), 147-169.
Dasgupta, Sakyashingha and Takayuki Osogami, 2017, “Nonlinear dynamic Boltzmann machines for time-series prediction,” Proceedings of the thirty-first AAAAI conference on artificial intelligence.
Hebb, Donald O., 1949, The organization of behavior, New York: Wiley & Sons.Hopfield, John J., 1982, “Neural networks and physical systems with emergent collective computational properties, “ Proceedings of the
National Academies of Sciences, 79, 2554-2558.Marks, T., and J. movellan, 2001, “Diffusion networks , products of experts, and factor anlaysis,” Proceedings of the Third International
conference on Independent compennet analysis and blind source separation.Osogami, Takayuki and M. Otsuka, 2015, “Learning dynamic Boltzmann machines with spike-timing dependent plasticity,” Techicial report
RT0967, IBM Research.
Boltzmann machines
34
Non-Bayesian
(Parametric)
Bayesian
(Non-Parametric)
Spatial Data
(Cross Sectional)
Sequential Data
(Time Series)
“Big Picture” for machine intelligence– four schools of thought
Gaussian
Process
Kernel Regression
Linear Smoothers
Linear Regression
Bayesian Linear Regression
Generalized
Linear Model
Neural Networks
CNNRNN
LSTM
Linear Time Regression
MAAR
ARMA
ARCH
GARCH
G-DyBM
DyBM
State
Space
ModelDynamic
Linear Model
Kalman Filter
UKF
EKF
GPK
GPTS
ARGP
Extend
Influence
Related
Rival
Connections
35
Machine intelligence’s fragility
36
What is Meant by “Fragility of Machine Intelligence”?
Fragility of Machine Intelligence: When a machine intelligence-based solution fails to meet accuracy or performance criteria expected of the model’s application, the model outcome may adversely affect dependent systems. The unexpected output of the machine intelligence is a result of the model’s fragility, or inability to process data in relation to the expected performance criteria.
Sources of fragility• Overlaying new systems on existing systems that are not likely to be seamlessly interoperable• Mismatching systems/algorithms with particular use cases• Interaction of humans and machines• Poorly designed system architectures• Poorly designed algorithms buried in a system architecture– algorithmic malpractice• Overall exposure to cyber attacks
37
Major vulnerabilities exist across many modern machine intelligence methods
38
Vulnerabilities Problems Typical Algorithms
“Blackbox” Difficulty in debugging, error tracking Neural Networks
Computational
Cost
1. Lack of computing power
2. High computational complexity (e.g. 𝑂(𝑛3))Neural Networks, RNN, Search Algorithms
Data
1. Biased/volatile/fat-tailed/arbitrary data
2. Requirement of well-labeled data
3. Insufficient data
Neural Networks, Clustering, Classification
Algorithms (RNN, Random Forest, SVM),
Assembly Algorithms
Security1. Cyber attacks
2. Privacy gets threatedNeural Networks, IoT
Algorithm
Design
1. Impractical vocabulary bank in NLP
2. Complex hyperparameter selection
3. Not scalable
4. Cannot solve non-linear problems
5. Not interpretable
6. Fail to capture error in learning process
Neural Networks, NLP, RNN, Gaussian
Regression, Kalman Filter, Sorting
Algorithms, Blockchain Algorithms,
Robotics Algorithms
Information
Loss
1. Memory loss in training process
2. Feature loss in dimensionality reduction
Neural Networks, PCA, Clustering
Algorithms
Regulation Lack of standardized regulations IoT
Impacts of Machine Intelligence Fragility
• Regulatory
• Impact to Peripheral Machine Intelligence Modules
• Impact to Transaction Flows
• Financial Impact to Business (resulting from Impact to Transaction Flows)
• Financial Impact to Business (as a “Residual Result”)
• Property Damage
• Human Injury or Loss of Life
39
Matching algorithm/system/process/ data to risk-related use cases
40
• Objectives/use cases:
– Identify trends (first moment estimation e.g., expected return)
– Identify short-term risk (second moment estimation e.g., volatility estimation)
– Identify long-term "risk" (higher moments, e.g., expected tail loss or expected shortfall from a skewed, non-normal [likely non-ergodic] distribution)
• OLS/Factor modeling (Explicit & implicit)
• Better regression techniques: Random forests, lasso regression
• ARIMA/GARCH/Vector Auto Regression (VAR)
• Recursive Neural Networks (RNNs)
• Boltzmann machines
– Restricted Boltzmann machines
– Dynamic Boltzmann machines
– Gaussian dynamic Boltzmann machines
• Gradient boosting
– XGBoost
– LightGBM
Matching techniques to objectives
41
Modeling challenges
• Variable data availability & quality
• Changing underlying data generation processes
• Moving up the ladder of causation (Pearl, 2018, p. 28)
– Level 1: Association (Seeing, observing)
– Level 2: Intervention (Doing, intervening)
– Level 3: Counterfactuals (Imagining, retrospection, understanding)
• Macroeconomic interaction
• Insurance market interaction (Game theory)
• Behavioral economics (Changing product design, marketing, distribution, strategy can change opportunities within a given segment)
• Digitizing society impact
42
How to persuade an individual to make a decision(Inspired by Koomey (2001) figure 19.1 p. 88)
Agree on objectives and criteria for acceptance
Disagree on objectives and/orcriteria for acceptance
Agree on data Computational decision Negotiate
Disagree on data Experiment Paralysis or chaos
43
Ensemble modeling is the process by which multiple weaker systems are combined to create more complex systems
Lay
er 1
Lay
er 2
Lay
er 3
Ou
tpu
t
44
Example: highly complex datasets, such as “donut problem” can’t be learned without ensemble modeling (Source: XLP)
B
A
B
A
Linear Model Non-Linear Model
When splitting the donut into binary classifications, a linear
model would simply cut the data in the middle. This leads
to a very high error rate of 50%, and is no better than
random guessing
A single, non-linear model would draw a curve that most
minimizes the error of binary classification. The error rate
using a non-stacked model would still be extremely high
45
Stacking to generate more complex system, ensemble models learn from data with greater accuracy (Source: XLP)
BA
Fir
st L
evel
Alg
ori
thm
Sec
on
d L
evel
Alg
ori
thm
New Prediction
y1
y2
y3
Ensemble models study the
errors of weaker tests,
creating more complex
systems with higher predictive
accuracy
46
Gradient Boosting algorithm has a high prediction accuracy using the original VIX data
47
0
20
40
60
80
100
120
140
160
180
200
220
1/1/1990 1/1/1995 1/1/2000 1/1/2005 1/1/2010 1/1/2015 1/1/2020
Historical AAPL Price
Predicted Values by LightGBM
LightGBM Predictions based on Historical VIX Data
Predictions and
real values almost
align as MSE =
0.00000106
Weaknesses of Existing ML/AI Tools
However, a minor jump in VIX can alter the accuracy of the model predictions significantly
48
0
20
40
60
80
100
120
140
160
180
200
220
1/1/20051/1/1990 1/1/1995 1/1/2000 1/1/2010 1/1/2015 1/1/2020
Historical AAPL Price
Predicted Values by LightGBM
LightGBM Predictions based on Altered VIX Data
Relatively larger
residuals are seen
as MSE = 9.229
Weaknesses of Existing ML/AI Tools
Gradient Boosting fails when data is limited, as the algorithm merely tries to find the similar pattern in the past data
49
0
1
2
3
4
5
6
7
8
9
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021
Predicted Values by LightGBM
Historical Data
Predicted Values by Linear Regression
Predictions by Linear Regression vs LightGBM
13 data
points in
total
A vanilla linear
regression manages to
capture the up trend
Booster acts as a classification tree here.
When new data come, it is only able to find
a similar pattern in historical data and then
assign a same value. In this case, booster
believes that the 3 data points in test data
are similar to the last data point in training
data. Therefore, the predicted values are
always same as the last value in historical
line. That fails to capture the overall trend
Weaknesses of Existing ML/AI Tools
Gradient Boosting fails when a proper window size is not well-selected for a cyclical time series
992
993
994
995
996
997
998
999
1,000
1,001
1,002
1,003
1,004
1,005
0 1 2 3 4 5 6 7 8 9 10 11
LightGBM with 3-Day Back Window
LightGBM with 4-Day Back Window
Historical Data
Predictions by LightGBM with
Different Window Size
A 3999-row
cyclical data,
with 3 data
points in a cycle
Weaknesses of Existing ML/AI Tools
50
Final remarks
51
52
Defining a research agenda at the intersection of machine intelligence and digitizing societies
• Focus more on…
– Addressing the challenge of data curation
– Communicating actionable insight
– Testing algorithms on real, useful data at scale
– Discussing how machine intelligence should integrate into a digitizing society (define algorithmic malpractice; shape regulatoryenvironment regarding data & machine intelligence)
• Focus less on…
– Developing more sophisticated algorithms without a specific applied use case
– Testing algorithms on toy datasets
53
Challenges
Collecting & curating suitable & sufficient data
Matching the type of machine intelligence with an objective or use case
Making algorithms interpretable and diagnosable
Defining what constitutes algorithmic malpractice in the machine intelligence arena
Dealing with data sparsity and non-stationary data-generating processes
Architecting and complying with data privacy regulations
Addressing conflicts arising from human notions of law, fairness, and justice and machine-intelligence capabilities that can circumvent protections
Addressing system fragility as interconnected and complex networks are infused with machine intelligence
Addressing growing cyber-risk in digital ecosystems infused with machine intelligence
Inspiration from biology (Theodosius Dobzhansky)
54
Nothing in a digitizing society makes sense except in the light of the evolution of
machine intelligence
55
Legal notice
©2018 Swiss Re. All rights reserved. You are not permitted to create any modifications or derivative works of this presentation or to use it for commercial or other public purposes without the prior written permission of Swiss Re.
The information and opinions contained in the presentation are provided as at the date of the presentation and are subject to change without notice. Although the information used was taken from reliable sources, Swiss Re does not accept any responsibility for the accuracy or comprehensiveness of the details given. All liability for the accuracy and completeness thereof or for any damage or loss resulting from the use of the information contained in this presentation is expressly excluded. Under no circumstances shall Swiss Re or its Group companies be liable for any financial or consequential loss relating to this presentation.
56