Artificial Intelligence: Artificial Neural Networks
Download as PPT, PDF11 likes8,208 views
This document summarizes artificial neural networks (ANN), which were inspired by biological neural networks in the human brain. ANNs consist of interconnected computational units that emulate neurons and pass signals to other units through connections with variable weights. ANNs are arranged in layers and learn by modifying the weights between units based on input and output data to minimize error. Common ANN algorithms include backpropagation for supervised learning to predict outputs from inputs.
Artificial Intelligence: Artificial Neural Networks
- 1. 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 + - - + + - +
- 2. 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
- 3. 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
- 4. 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
- 5. 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
- 6. 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)
- 7. Basic Representation Sum: Netj = Sum of Wji Xi Transfer: Yj = F (Netj ) S u m Transfer X1 X2 X3 jth Computational Unit Weights Wj1 Wj2 Wj3 Yj Output Path
- 8. ANN • Computational units in ANN are arranged in layers - input, output, and hidden layers • Units in a hidden layer are called hidden units
- 9. 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
- 13. ANN Learning Algorithm Supervised Learning Unsupervised Learning Binary Input Continued Binary Continued Hopfield Net Perceptron ART I ART II Boltzman- Backpropagation Self-organizing Machine (popular algorithm widely used) Map
- 14. 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
- 15. Backpropagation • The units on a layer have full connections to units on the adjacent layers, but no connection to units on the same layer
- 16. 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
- 17. 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
- 18. Five Input Var. Net Working Capital/Total Assets Retained Earning/Total Assets EBIT/Total Assets Market Value of Common and 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
- 19. 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
- 20. Properties of ANN • No special recovery mechanism is required for incomplete information • Learning capability
- 21. Disadvantages of ANN • Black box – Difficulty to interpret information on the network • Complicated Algorithms