Title: NEURAL NETWORKS
1NEURAL NETWORKS
2Introduction
- The branch of AI known as "neural networks" or
"artificial neural networks" has become an
increasingly important area of AI. - Neural networks approaches building intelligent
systems using architectures and processing
capabilities that mimic some biological processes
of the human brain.
3Basics
- Biological system
- Brains are composed of neurons, which have the
unique characteristic that of all types of cells
in the body, they do not die. - biological neural networks
- http//psych.hanover.edu/Krantz/neurotut.html
- good basic tutorial and introduction
4Portion of a Network Two Interconnected
Biological Cells
5Biological system
- Human brains are estimated to contain up to 100
billion neurons (1011) and hundreds of different
types. - Neurons are found in groups called networks (now
you know why this area of AI is called neural
networks), each containing thousands of highly
interconnected neurons.
6Biological system
- dendrites - takes in information, provide inputs
to the cell, note they have plenty of surface
area to facilitate connection to axons of other
cells - axon protuberance that delivers outputs from
the neuron to connections with other neurons
7Biological system
- A neuron does nothing until the collective
influence of all its inputs reaches a threshold
level. - At that point, the neuron produces a
full-strength output in the form of a narrow
pulse that proceeds from the cell body, down the
axon, and into the axons branches. - it fires Since it fires or does nothing it is
considered an all or nothing device.
8Biological system
- synapse gap between the axon and dendrites.
- Stimulation at some synapses will encourage the
neuron to fire, while stimulation at others
discourages the neuron from firing. - Increases or decreases the strength of connection
and causes excitation or inhibition of a
subsequent neuron
9Artificial Systems
- Simulated neurons are viewed as a node connected
to other nodes via links that correspond to
axon-synapse-dendrite connections. - Each link is associated with a weight.
- The weight determines the nature (/-) and
strength of the nodes influence on another. - If the influence of all the links is strong
enough the node is activate (similar to the
firing of a neuron).
10Processing Information in an Artificial Neuron
11Artificial Systems (continued)
- Processing element
- Think of a PE as an artificial neuron.
- Receives inputs, processes the outputs, and
delivers a single output. - Inputs can be received as raw data or from
another PE
12Artificial Systems (continued)
- Network
- Composed of a collection of PE's grouped in
layers. - This example has three layers, the middle layer
is referred to as the hidden layer. - Structure
- Several possible structures
13Artificial Neural Network with One Hidden Layer
14Processing Information in an Artificial Neural
Network
- Inputs
- Each input is a value of a single attribute.
- If I wanted to predict stock prices, one
attribute of interest might be "volume" and
therefore, I would input the volume (number of
shares sold) on a specific day as one input.
15Processing Information in an Artificial Neuron
16Processing Information in a Network
- Outputs
- Solution to the problem. For instance, the
projected price of the stock
17Processing Information in an Artificial Neural
Network
- Weights
- Used to express the relative strength of an input
value or from a connecting PE (i.e., in another
layer). - These weights are essential, it is by adjusting
these weights that a neural network learns.
18Processing Information in a Network
- Summation function
- Used to compute a single value (weighted average)
from all the inputs to a particular PE. - Think of it as the internal stimulation or
activation level of the neuron
19Processing Information in an Artificial Neural
Network
- Transformation (Transfer) Function
- based on the summation value,
- the value, the transformation (transfer) function
produces an output.
20Summation Function for Single Neuron(a) and
Several Neurons(b)
21Processing Information in a Network
- There are many possible transformation functions,
the sigmoid function is popular. - Sometimes a threshold value is used, which is
easier to explain and understand.
22Processing Information in a Network
- For instance, for summation values less than .5 a
0 might be output, for summation values greater
than or equal to .5 a 1 might be output
23Learning Training in Neural Networks
- Neural networks are trained using data referred
to as a training set. - The process is one of computing outputs, compare
outputs with desired answers, adjust weights and
repeat.
24Learning Training in Neural Networks
- It is necessary to have a fairly large training
set, and you need to have the answer for each
case in the training set. - Discrepancies between the "right answer" (from
the training set) and the computed answer are
measured and based on the error, adjustments made.
25History of Neural Networks
- Basic research on brains dates back quite far.
- 1791 - Luigi Galvani (from Bologna) stimulated a
frog's muscles with electricity, leading to the
discovery that the brain has electrical activity - 1837 - Gocli observed the structure of neurons
with axons and connections to dendrites
26History of Neural Networks
- 1887 - Sherrington Synaptic interconnection
suggested - 1920's - discovered that neurons communicate via
chemical impulses called neurotransmitters. - 1930's - research on the chemical processes that
produce the electrical impulses.
27History of Neural Networks
- 1943 - McCullock and Pitts showed that a NN could
be used to code logical relationships such as - "x AND y" or "x OR y"
- 1950's - Hodgkin and Huxley were awarded the
Nobel Prize for work developing the model and
recording the electrical signal of the brain at
the cellular level
28History of Neural Networks
- 1969 - Minsky and Papert wrote Perceptrons
- showed that one-layer neural networks could not
handle statements such as - (x AND NOT y) OR (y AND NOT x)
- Based on this finding, they conjectured that
multi-level NN's would not perform better - Result funding for NN research dried up, for
about 10 years
29History of Neural Networks
- 1987 - Robert Hecht-Nielsen mathematically
disproved Minsky's and Papert's conjecture
regarding multi-layer neural networks not being
able to perform better than one-layer neural
networks. - Since then, this area has been subject to more
research.
30Basic Network Structures
- associative - single layer is representative
- hidden layer - can have more than one hidden
layer, note that it is uni-directional - double-layer - feeds forward and backward,
develops its own categories for representing the
data
31Neural Network Structures
32Artificial Neural Network Develop-ment Process
Get More, Better Data
Refine Structure
Select Another Algorithm
Reset
Reset
33Developing Neural Networks
- Step 1
- collect data
- Step 2
- separate data into training and test sets,
usually random separation - ensure that application is amenable to a NN
approach
34Developing Neural Networks
- Step 3
- define a network structure
- Step 4
- select a learning algorithm
- affected by the available tools shells available
- Step 5
- set parameter values
- affects the length of the training period
35Developing Neural Networks
- Step 6
- transform Data to Network Inputs
- data must be NUMERIC, may need to preprocess the
data, e.g., normalize values for a range of 0 to
1 - Step 7
- start training
- determine and revise weights, check points
- Step 8
- stop and test iterative process
36Developing Neural Networks
- Step 9
- implementation
- stable weights obtained
- begin using the system
37Example - Financial Market Analysis
- Karl Bergerson of Neural Trading Co uses Neural,
a trading systems with BrainMaker and a C-based
E.S. for money-management rules. - Using 9 years of hand-picked financial data,
trained the NN and ran it against a theoretical
10,000 investment. - After 2 years, the fictional account had grown to
76,034 (660 appreciation).
38Financial Market Analysis - continued
- When tested on new data, 89 accurate.
- Developer quoted "Neural nets are the best tools
for pattern recognition, but you can't just dump
data into one and expect to get wonderful
results. The most important factor is your
training data. You have to have your whole act
together, training, design, and the right tools.
39Financial Market Analysis - continued
- Some of the attributes used price, volume,
advance/decline etc. - The neural network predicts market fluctuation
and the expert system component flags buying or
selling opportunities
40Sales Support
- Veratex Corp. distributes medical and dental
products. - They send unsolicited catalogs to physicians and
dentists. - When a customer buys something, their name is
added to the customer database. - 40 telemarketers then call the names in the
database for reorders
41Sales Support - continued
- The problem
- many dormant accounts, i.e., customers who had
not placed reorders. - The telemarketers are not trained to prospect for
new clients and they only have about 20 of their
time allocated for calling dormant accounts.
42Sales Support - The Problem (continued)
- The database contains 44,000 customers, these
represent potential business that is not being
tapped. - Further, as the data ages, it becomes less
reliable (i.e., physicians and dentists move and
retire).
43Sales Support - continued
- The solution
- The company hired Churchill Systems to build a
back-propogation (a learning algorithm) to
identify those customers in the dormant pool most
likely to place reorders. - With this information, telemarketers could focus
their limited time on customers with the most
potential.
44Sales Support - continued
- System was built using NNU400 neural network
utility (from IBM). - Inputs consisted of statistical and demographic
data culled from Dun Bradstreet and other
sources. - The network was applied against the customer
list, giving each customer a numerical rating
which was put into the customer records and then
used as a sort key.
45Sales Support
- Results
- President of Churchill Systems "More Veratex
accounts were reopened in five months, than
similar periods (without the network). - "The patterns and
- interrelationships uncovered by the neural
network proved to be an extremely valuable
resource for Veratex marketing analysts."
46Sales Support - continued
- General comments
- "A lot of people think you can avoid knowledge
engineering. Forget it - you can't do it. You
really have to get down to the business problem
before you can do anything else. - In fact, Light claims that a large part of the
neural network's development time entailed
gathering, cleaning up, and organizing the
appropriate data.
47Horse Bloodlines
- University of California at Davis School of
Veterinary Medicine conducts blood tests to
confirm the bloodlines of Thoroughbred horses. - Thoroughbreds cannot be raced unless their
bloodlines are known. - To do this, 142 separate reaction tests must be
run on a blood sample. - as many as 72,000 tests per day.
48Horse Bloodlines - continued
- The problem how to automate this function so
that a technician didn't have to perform this
job. - A neural network was trained to read these tests
starting April 1987, was pilot tested for about
one year (1989?, article printed in 1990).
49Horse Bloodlines - continued
- The neural network was trained to read a blood
test and determine if a reaction occurred - a
simple yes or no. - The neural network must be accurate, the results
of the lab cannot be questioned or breeders will
not use the lab.
50Horse Bloodlines - continued
- Specifically, the network must "look" at a drop
of blood and decide if the cells in the drop of
blood have clumped together (agglutinated). - The system "sees" using a video camera that
divides the field of vision into 262,144 pixels
of information.
51Horse Bloodlines - continued
- Using this information, the developer believed it
would have taken 28 million years to teach the
network the concept of "clumpiness". - It was just too much raw data.
52Horse Bloodlines - continued
- Lendaris (1970) pioneered the scanning of aerial
surveillance photographs by computer to detect
orderly man-made features such as orchards, road
intersections, etc. - The contribution that could be applied to the
blood testing problem was the use of the Fourier
transform, developed by a 19th-century French
physicist and mathematician.
53Horse Bloodlines - continued
- This transformation converts massive amounts of
data into oscillating waves of energy and can be
used to highlight sharp gradations, such as the
edge of a building or the edge of a clump of
blood. - The Fourier transformed 262,144 pixels into 48
data points that the network was easily trained
to recognize.
54Horse Bloodlines - continued
- The neural network tool used was supplied by
Science Application International Corp (of San
Diego), cost 25,000 and called Delta II. - It includes both software and an accelerator
board to enhance a 386 machine.
55Horse Bloodlines - continued
- The developer of the system to read blood tests
is skeptical about the commercial development of
Neural Nets - "The difficulty is, what will you sell? A neural
net is just an algorithm - a method of
calculation like a statistical regression or
multiplication. It is hard to protect a product
like that - hard to get a commercial handle on
it."
56Horse Bloodlines - continued
- He believes that what will succeed are a variety
of hardware systems with the neural network
learning method automated and embedded in the
hardware. - This will also alleviate the user from having to
understand as much about the neural network.
57KBS vs. Neural Networks
58(No Transcript)
59Advantages of Neural Nets
- Able to learn any complex non-linear mapping (31)
- Do not make a priori assumptions about the
distribution of the data/input-output mapping
function (30) - Very flexible with respect to incomplete,
missing, noisy data, fault tolerant (29) - Easily updated, suitable to dynamic environments
(15) - Overcome some limitations of other statistical
methods, while generalizing them (15) - Hidden nodes, in feed-forward, can be regarded as
latent/unobservable variables (5) - Can implement on parallel hardware, increasing
accuracy and learning speed (4) - Can be highly automated, minimizing human
involvement (3) - Specially suited to tackle problems in
non-conservative domains (3)
60Disadvantages of Neural Nets
- Lack theoretical background, no explanation,
black box (28) - Selection of network topology and parameters
lacks theoretical background, trial and error
(21) - Learning process can be very time consuming (11)
- Can overfit the training data, becoming useless
for generalization (10) - No explicit set of rules to select a suitable ANN
paradigm/learning algorithm (8) - Too dependent on the quality/ amount of data
available (6) - Can get stuck in local optima, narrow valleys
during training (5) - Techniques still rapidly evolving and not
reliable or robust enough yet (3) - Lack classical statistical properties.
Confidence intervals and hypothesis testing are
not available (2)
61Potential Problems with Neural Networks
- The military has been experimenting with ANN
techniques for sometime. - One application of interest was to identify
objects on the battlefield at night. For
instance, distinguishing the difference between a
tank and a rock. - A scanner showed an automated neural network
thousands of photographs of tanks, rocks, and
other battlefield objects.
62Potential Problems with Neural Networks (cont.)
- After training the ANN could correctly
distinguish a tank from a rock 100 of the time. - Later, it was discovered that all the photos of
the tanks had been taken with the same camera. - The tank photos were all slightly darker than the
photos of the other objects. - What the ANN had really learned was to identify
the camera used to take the picture, not the
difference between rocks and tanks!