CSC352 Final Project Proposal

1
CSC352 Final Project Proposal

• Melissa Patton and Victoria Manfredi
• Smith College
• November 20, 2000

2
Introduction and Outline

• Were going to present the following
• Possible Projects
• Neural Networks
• Time-line
• Conclusion

3
Possible Projects

• Exploring NESL. NESL is a free, nested
  data-parallel programming language that works on
  Unix.
• Implementation of Parallel Graph Algorithms on
  the MasPar- a journal article. MasPar Parallel
  Language (MPL) was used.
• Neural Networks

4
Neural Networks Outline

• Reasons for Choosing
• Background Info
• Specifics
• Proposed Implementation
• Area of Focus
• Limitations and Assumptions
• Things to Figure out

5
Neural Networks - Reasons for Choosing

• When doing a web search on parallel algorithms,
  results about neural networks kept showing up
• Chose to use XOR neural network for its simplicity

6
Neural Networks - Background Info

• Based on the nervous system
• Graph with weighted edges
• Takes some input and produces a certain output
• Neural nets trained by learning algorithms
• Example uses optical neurocomputer, credit
  card screening, target advertising

Resources for background info
http//sunsite.univie.ac.at/textbooks/statisitics/
stneunet.html
and http//web.singnet.com.sg/midaz/I
ntronn.htm
7
Neural Networks Outline

• Reasons for Choosing
• Background Info
• Specifics
• Proposed Implementation
• Area of Focus
• Limitations and Assumptions
• Things to Figure out

8
Neural Networks - Specifics

• Currently plan to use SNNS, a Unix-based, free,
  neural network simulator with a graphical
  interface, and the Parallaxis programming
  language.

Figure 1. Example of SNNS interface

• The XOR neural net will be used.

9
Neural Networks - Proposed Implementation

• Load XOR neural net into a simulator and train it
  until output correct.
• Input weights from simulator and XOR neural net
  into Parallaxis program.
• Try to have the Parallaxis program also give the
  correct output.
• If this works, look at speedup (compare parallel
  and sequential implementations)

Figure 2. Example of XOR neural net
10
Neural Networks - Area of Focus

• Demonstration that it can be done. Will the
  Parallaxis program produce the same output as the
  neural net simulator?

11
Neural Networks Outline

• Reasons for Choosing
• Background Info
• Specifics
• Proposed Implementation
• Area of Focus
• Limitations and Assumptions
• Things to Figure out

12
Neural Networks - Limitations and Assumptions

• Limitations Parallaxis program will not be able
  to train a neural net (although, if we have time
  it might) and it will only work for XOR neural
  net
• Assumptions Parallaxis will allow PEs to be
  arranged into XOR configuration

13
Things to Figure Out

• SNNS - How to work with SNNS, how to input a
  neural net, how to train the net, and how to get
  the actual weights from a trained neural net
• How to relate nodes and edges of neural net to PE
  configuration. Want the adjacency list
  implementation for a graph, using the weight
  instead of 1 or 0.

14
Introduction and Outline

• Were going to present the following
• Possible Projects
• Neural Networks
• Time-line
• Conclusion

15
Time-line

• Nov. 27 - Have SNNS and how to obtain the weights
  from neural nets in it figured out
• Dec. 4 - Have figured out how to map XOR neural
  net to PE configuration, and have done some
  serious work on the Parallaxis program
• Dec. 11 - Have programming part of project done,
  and most or all of work done on Report and Poster

16
Conclusion

• In summary, what we want to do
• -Use neural network simulator to get weights for
  
• XOR neural net
• -Input XOR neural net and weights into
• Parallaxis program
• -Have Parallaxis programa give the correct output