www.vintasystems.com An Introduction to Neural Networks Welcome! Welcome!
Transcript
PowerPoint PresentationWHAT ARE NEURAL NETS?
Outputs
Inputs
Il
Ii
Ix
OI
Oj
Oy
Neural Nets learn the relationship between the independent variable
and the dependent variable
WHAT ARE NEURAL NETS?
Neural Nets are function approximation tools
Neural Nets learn the relationship between the independent variable
and the dependent variable
WHAT ARE NEURAL NETS?
NEURAL NETS vs STATISTICAL APPROACHES
Neural Nets make no assumptions on the statistical distribution of
data
Neural Nets are non-linear approaches that provide much accuracy
when modeling complex data patterns
Logistic Regression is one Neuron Neural Net
CREDIT RISK MODELING
Current Banking Practice
Use regression techniques to derive a model relating the input risk
factors and a credit decision
The estimated coefficients in the Regression Model are used to
derive the weights in a credit scorecard
Use historical data for model derivation
Modern AI-Based Practice
Use AI-Based Model (Neural Net, Rules, Hybrid) to derive a model
relating the input risk factors and a credit decision
AI-Based Model can detect complex patterns in data that humans are
unable to see. Also uses Historical data
NEURAL NET BUILDING BLOCK
A neuron receives inputs via weighted links from other
neurons
The inputs are processed according to the activation function of
the neuron
The outputs of a neuron are passed to other neurons in the neural
net
Outputs
Inputs
xi
yi
Neuron
Output
Hidden
Input
Input Neuron – receives encoded information from the external
environment
Hidden Neuron – allows intermediate calculation between inputs and
outputs
Output Neuron – Sends signals to the external environment in the
form of an encoded answer to the problem presented
Best known Neural Net for business application is the Multi-layered
Feed Forward Neural Net (MFNN) with the Back Propagation Learning
Rule
95% of the reported Neural Net Business Applications utilize
MFNN
TYPES OF NEURONS IN A NEURAL NET
HOW DO YOU TRAIN A NEURAL NET?
Neural Nets learns from examples
Collection of examples is known as the training set.
Training set must be prepared properly. Good examples must be
used.
Use as much relevant input as possible. Training set must be a
representative set.
Remove outliers in the data.
As examples from the training set are fed in the neural net, the
weights interconnecting the neurons evolve (learning
process).
Objective of the learning process is to minimize the training error
(difference between the neural net output and the desired output
supplied in the training set).
Weight
Output
Neurons
Link
MFNN Learning
Each input pattern (Example: (xi…xn)) produce an output Ok
Each Neuron calculates output yj
as follows:
logistic function
Ok = f (Net Input)
HOW DOES A NEURAL NET LEARN?
The output Ok is compared to the desired output Dk found on the
training set.
Based on the discrepancy between the Neural net Output Ok and the
desired output Dk , the neural net weights are updated.
Weight update rules that minimizes the total average-squared error
(or SSE) are used.
For Back propagation learning, minimization techniques in calculus
are used to derive the weight update rules.
Convergence is achieved when calculated SSE is within the tolerance
limit.
SAMPLE NEURAL NET TRAINING
To approximate “AND” function
Calculate Net Input
NEURAL NET MODELING STEPS
Representative Data Acquisition
Data Remapping / Transformation
Removal of Outliers
MODEL GENERATION
MODEL VALIDATION
MODEL RETRAINING
ACTUAL NEURAL NET APPLICATION IN BANKING & FINANCE
VISA International implemented an operational fraud detection
system based on a neural net
HSBC (Falcon Project) used neural nets for credit card fraud
detection
Money transfers between New York and London banks are monitored to
detect possible money launderers using Neural Nets & other
AI-Based Models.
A lot of big banks worldwide have adopted Neural Nets as their
underlying technique in making credit decisions.
A lot of banks have used Neural Nets for validating bank
signatures.
Finance companies have used neural nets for exchange rate
forecasting, derivative security pricing, future price forecasting,
bankruptcy prediction and stock performances and selection
prediction.
Banks have used neural nets for Customer Relationship
Management.
27-C Rufino Pacific Tower
Makati City