Artificial Intelligence: Artificial Neural Networks
Post on 21-Nov-2014
334 Views
Preview:
DESCRIPTION
Transcript
Artificial Neural Networks (ANN)• Human information processing takes place through the interaction of many
billions of neurons connected to each other, each sending excitatory or inhibitory signals to other neurons (excite in positive/suppress in negative)
• Human Brain: Parallel Processing
+ excites- supresses
+ -
+ +
- +
-
ANN
• The neuron receives signals from other neurons, collects the input signals, and transforms the collected input signal
• The single neuron then transmits the transformed signal to other neurons
ANN• The signals that pass through the junction, known as synapses, are
either weakened or strengthened depending upon the strength of the synaptic connection
• By modifying synaptic strengths, the human brain is able to store knowledge and thus allow certain inputs to result in specific output or behavior
• Translates into a mathematical model
• Artificial Neural Networks compare weights– Synopsis is small = -– Synopsis is large = +
• ON = +
• OFF = -
• Neurons are trained– Neurons are on (+) or off (-)
• Example: Could be Facial Recognition
ANN• A basic ANN model consists of
– Computational units – Links
• A unit emulate the functions of a neuron• Computational units are connected by links with
variable weights which represent synapses in the biological model (Human Brain)
• Learning Curve: Change synopsis in face recognition– Changes & learns new info
ANN
• The unit receives a weighted sum of all its input via connections and computes its own output value using its own output function
• The output value is then propagated to many other units via connection between units
Basic Representation• Parallel Transfer
– Some connections bi-directional, some one-way
• Variation of algorithms
– 2 levels
– Multi-levels
• y=f (x1, x2, x3)
– where is is a transform function (linear or non-linear)
Basic Representation
TransferSum
X1
X2
X3 jth ComputationalUnit
Weights
Wj1
Wj2
Wj3
Yj
Output Path
Sum: Netj = Sum of Wji Xi
Transfer: Yj = F (Netj )
ANN
• Computational units in ANN are arranged in layers - input, output, and hidden layers
• Units in a hidden layer are called hidden units
Hidden Units
• Hidden unit is a unit which represents neither input nor output variables
• It is used to support the required function from input to output
ANN Learning Algorithm
Supervised Learning Unsupervised Learning
Binary Input Continued Binary Continued
Hopfield Net Perceptron ART I ART IIBoltzman- Backpropagation Self-organizing Machine (popular algorithm widely used) Map
Backpropagation• The algorithm is a learning rule which
suggests a way of modifying weights to represent a function from input to output
• The network architecture is a feedforward network where computational units are structured in a multi-layered network: an input layer, one or more hidden layer(s), and an output layer
Backpropagation
• The units on a layer have full connections to units on the adjacent layers, but no connection to units on the same layer
Backpropagation
• Calculate the difference (error) between the expected and actual output value
• Adjust the weights in order to minimize the error
• Minimize the error by performing a gradient decent on the error surface
Backpropagation
• The amount of the weight change for each input pattern in an epoch is proportional to the error
• An epoch is completed after the network sees all of the input and output pairs
Five Input Var.
Net Working Capital/Total Assets
Retained Earning/Total Assets
EBIT/Total Assets
Market Value of Commonand Preferred Stock/Book Value of Debt
Sales/Total Assets
Two Output Variables
Solvent Firms
Bankrupt Firms
An ANN model to Predict a Firm’s Bankruptcy
Advantages of ANN
• Parallel Processing
• Generalization– a great deal of noise and randomness can
be tolerated
• Fault tolerance– damage to a few units and weights may
not be fatal to the overall network performance
Properties of ANN
• No special recovery mechanism is required for incomplete information
• Learning capability
Disadvantages of ANN
• Black box– Difficulty to interpret information on the
network
• Complicated Algorithms
top related