Archaeological Auto Classification System

1
Archaeological Auto Classification System
2
History of Development

• Spring 2002 Concepts of Data Management collides
  with Information Systems.
• Skip Lohse and Corey Schou discuss automation of
  classification systems.
• Summer 2002 Corey Schou Proposes having the
  Advanced System Analysis and Design class build
  an Expert System for Skip.
• Fall 2002 The Advanced System Analysis and
  Design class was given the assignment of
  developing an application to classify Projectile
  Points.

3
Development Team
4
Development Strategy

• Identify all the potential classification Types
• Identify the Classification Method
• Generate an Image Manipulation system from the
  Image Pro Plus Image Library
• Identify the information needed from each image
• Develop a Database to hold the information
  gathered form the Images

5
Results of Initial Development

• 18 types were identified.
• Discriminant Analysis was dropped in favor of
  Logistic Regression.
• Decided to write our own Image Manipulation code
  to avoid Licensing Issues.
• Prototype Database in MS Access, then upsize to
  MS SQL server or other Industrial Database.
• Eventually dropped Logistic Regression in favor
  of Artificial Neural Networks due to problems
  getting the computer to measure an image.

6
Initial Required Measurements
7
What is an Artificial Neural Network?

• ANNs are an attempt to model the way a brain
  works by creating a series of artificial Neurons,
  called Nodes, that are tightly connected just
  like a biological brain.

8
Comparison
9
ANNs Continued

• Just like real Neurons, Nodes can have two
  states
• Activated
• Un-Activated
• Their state depends on the results of an
  Activation Function.

10
Activation Functions

• There are two types of Activation functions
• Step Activation
• Sigmoid Activation

11
Comparison of Activation Functions
12
How does a Node Activate?

• The Node sums the values of all its inputs,
  applies the result to the Activation function,
  and then sends the resulting value through its
  outputs.
• The Input values are modified by a Weight
  factor of the path from the outputs of the
  previous nodes.

13
Inputs modified by Weights

• Outputs

• Hidden Layer

• Inputs

14
Results of the Neural Network

• As the initial input values travel through the
  ANN, they are modified until they reach the
  Output Layer.
• The various output values are examined, and the
  output with the largest value is the Result.
• The results of all the output nodes can be
  normalized so that their results fall between 0
  and 1, which gives a probability result for
  each node, rather than a single Result for the
  whole network.

15
Training a Neural Network

• Supervised
• Unsupervised
• We use Backpropagation

16
Backpropagation

• The main step of training on a pattern utilize
  the following steps.
• Present the pattern at the input layer.
• Let the hidden units evaluate their output using
  the pattern.
• Let the output units evaluate their output using
  the result in step 2) from the hidden units.
• The steps 1) - 3) are collectively known as the
  forward pass since information is flowing
  forward, in the natural sense, through the
  network.
• Apply the target pattern to the output layer.
• Calculate the error on the output nodes.
• Train each output node using gradient descent.
• For each hidden node, calculate its error.
• For each hidden node, use the error found in step
  7) to train according to gradient descent.
• Steps 4) - 8) are collectively known as the
  backward pass
• Step 7) involves propagating the error from those
  output nodes in the hidden unit's outputs back
  towards this node so that it can process them.
  This is where the name of the algorithm comes
  from.

17
Backpropagation
18
What are the inputs?

• An image of a point is taken and its outline is
  extracted.
• The outline is then split up into a set of
  smaller pieces.
• Those pieces are then reduced to Tokens, which
  are fed into the network.

19
The steps in generating inputs for the network
20
Demo of the Project

• AACS Project

21
Problems with the current version

• Various display problems
• Problems training the network
• The tokenizer was wrong
• The network training function was bad
• The training set is incomplete
• The training set lacks definition

22
Definition problems
Rabbit Island Stemmed
Mahkin Shouldered Lanceoalate
23
The Future

• Attach a database to the interface to capture the
  outputs.
• Modify the interface to allow for multiple
  artifact types and classification systems.
• Design a web interface to allow internet access.
• Two levels of availability Research and Public.

24

• The End