Topics in Artificial Intelligence

1
Topics in Artificial Intelligence

02.18/2008
2
Agenda
1
AI Overview
Artificial Neural Nets
Questions/Discussion
Questions/Discussion
3
What is Artificial Intelligence?

• Impossible to define

4
Okay, seriously.

• Advanced algorithms for complex problems
• Interdisciplinary
• Incubator of new disciplines

5
Some subfields

• Robotics
• Knowledge representation
• Planning
• Machine learning
• Natural Language Processing
• Perception
• Affective Computing

6
Strong AI

• Combining everything to mimic humans
• Ripe with philosophical contention
• The Turing test
• Searles Chinese Room
• Computational neuroscience

7
Artificial Neural Networks

• biologically inspired
• non-linear
• stochastic

8
History of Artificial NNets

• Originated in the late 50s
• Rosenblatt
• The Perceptron (1960)
• Diminished in the late 60s
• Minksy
• Perceptrons (Minksy and Papert, 1969)
• Re-emerged in the 80s
• Back-propagation

9
The Biological Neuron
10
The Biological Neuron
11
Synaptic connections

• Highly complex
• Placement of connection
• Direction of transmission
• Can be temporally staggered
• 1015 connections in humans
• Strengthen with reinforcement
• Key to learning

12
The Artificial Neuron

• X vector
• W vector
• Transfer function (u)
• Training function (t)

13
Transfer functions

• Sum of inputs comes in
• Output goes out
• input (?1mwixi) b
• u(input) y

14
Transfer functions

• Parameters
• a low value
• b high value
• c low threshold
• d high threshold
• Shapes
• Step
• Ramp
• Sigmoid

15
Training functions

• Supervised learning
• Relies on training cases
• Perceptron learning
• Error backpropagation (gradient descent)
• Unsupervised learning
• Relies on clustering or other patterns
• Reinforcement learning
• Relies on feedback from the environment

16
Simple Perceptron

• Step function
• a 0
• b 1
• c 0
• u(input) 0 if inputlt0
• 1 if inputgt0

17
Perceptron Training

• Training data
• X vector of test cases
• xi vector of inputs for test case i
• xki input k for test case i
• di desired output for test case i
• a coefficient of learning (0 lt a lt 1)
• Algorithm
• for each xi, di in X
• for each xki in xi
• wk wk a(di yi)xik

18
Perceptron Example

• Classification
• 2D space
• Linear separation
• Given a point, is it black? 0 or 1

19
Perceptron Example

• Setup
• Training set
• Random initial weights/bias
• w1 .01
• w2 .5
• b -.2
• Sufficient a .15

20
Perceptron Example

• Spreadsheet for math
• w1 .115
• w2 .695
• b .1
• More complex example
• online applet

21
Further topics

• N-dimensional perceptrons
• Multilayer perceptrons
• Hopfield Networks
• Error backpropagation

22
Thank You!
www.PlatinumSolutions.com