Reference no: EM132372096
Neural Networks
Introduction
For this assignment you will do tasks using the MLP in c# supplied to investigate some data sets.
You can get marks for task 4 without doing task 3 (each task is independently marked)
Task 1 - Train a neural network on the Cancer dataset so it's not over or under trained then fill out a short form/report. Completing this task only will give you up to a maximum of 1 mark and, I (and Angus) will be happy to help you with this in the tutorials.
Task 2 - Basically just train a neural network on the Card dataset so it's not over trained without help from me or anyone else then fill out a short form/report. This should be your own work (if you complete this and the previous task you are eligible for a mark up to 2.5).
Task 3 - is to train 2 neural networks on the Wine dataset. But use the 33-33-33 for the first one then 33-66 methodology for the second one. I will be interested in what you discover please give me some thoughts on whether this works.
Task 4 - is to train a neural network on the Weed seed dataset. This more of an exploration and is for students that are really interested, its very optional and this is a dataset I know less of so I will be interested in what you discover.
Note the datasets will need pre-processing.
Task 1 - Train a basic ANN so it's correctly trained
Using the Cancer dataset that I supply in the MT1.zip file, train a neural network using the 33-33-33 split into test, train, and validation data so it has:
• A sensible topology (number of hidden nodes);
• A sensible training constant (eta);
• A sensible number of training epochs
You must also estimate the accuracy of the resultant network
Filling out the following form is suitable as a one page submission for this task.
|
Name
|
|
|
Student Id
|
|
|
Data set picked
|
Cancer
|
|
Training method
|
Hidden Nodes tried:
Eta (training constants) tried:
|
|
Details of MLP
|
Hidden Nodes:
Epochs :
Eta (training constant):
Testing Accuracy :
Training Accuracy :
Validation Accuracy :
|
|
Best estimate of validation accuracy for a generalised solution.
|
|
|
Comments
|
|
Task 2 - Train a basic ANN so it's correctly trained
Using the Card dataset that I supplied, train a neural network using the 33-33-33 split into test, train, and validation data. After training it should have:
• A sensible topology (number of hidden nodes);
• A sensible training constant (eta);
• A sensible number of training epochs
You must also estimate the accuracy of the resultant network
Filling out the following form is suitable as a one page submission for this task.
|
Name
|
|
|
Student Id
|
|
|
Data set picked
|
Card
|
|
Training method
|
Hidden Nodes tried:
Eta (training constants) tried:
|
|
Details of MLP
|
Hidden Nodes:
Epochs :
Eta (training constant):
Testing Accuracy :
Training Accuracy :
Validation Accuracy :
|
|
Best estimate of validation accuracy for a generalised solution.
|
|
|
Comments
|
|
|
|
|
WARNING : The early version of normalise found on Data.cs has a bug that stops this data set from normalising - if you are using angus's or and early ANN that you modified please use the new Data.cs supplied or at least fix the bug line shown below (this may be irrelevant in 2019, at least in the v7 version).
double mul = 1 / (max - min);
normMul[i] = mul;
// Becomes :
double mul;
if (max - min > 0.0001) mul = 1 / (max - min); else mul = 1;
normMul[i] = mul;
Task 3 - Use ANN to explore a dataset
Using the Wine dataset that I supplied, train a neural network using the 33-33-33 Methodology;
• A sensible topology (number of hidden nodes) (including the 66-33 network);
• A sensible training constant (eta) (including the 66-33 network);
• A sensible number of training epochs (including the 66-33 network)
You must also estimate the accuracy of the resultant network -
Comment on your observations of how this changes things:
Filling out the following form is suitable as a one page submission for this task.
|
Name
|
|
|
Student Id
|
|
|
Data set picked
|
|
|
Details of MLP 1 (33-33-33)
|
Hidden Nodes: Epochs :
Random seed(s):
Eta (training constant):
Testing Accuracy :
Training Accuracy :
Validation Accuracy :
|
|
Details of MLP 2 (66-33)
|
Hidden Nodes: Epochs :
Random seed(s):
Eta (training constant):
Testing Accuracy :
Training Accuracy :
Validation Accuracy :
|
|
Accuracy for a generalised solution. 33-33-33
|
Average Accuracy:
|
|
Accuracy for a generalised solution. 66-33
|
Average Accuracy:
|
|
Comments about this exercise
(marks here)
|
|
|
General Comments
|
|
Task 4 - Train a basic ANN so it's correctly trained
Using the Weedseed dataset that I supplied, train a neural network using the 33-33-33 split into test, train, and validation data. After training it should have:
• A sensible topology (number of hidden nodes);
• A sensible training constant (eta);
• A sensible number of training epochs
You must also estimate the accuracy of the resultant network
Tell me what you learned about this data set if anything
Filling out the following form is suitable as a one page submission for this task.
|
Name
|
|
|
Student Id
|
|
|
Data set picked
|
|
|
Training method
|
Hidden Nodes tried:
Eta (training constants) tried:
|
|
Details of MLP
|
Hidden Nodes:
Epochs :
Eta (training constant):
Testing Accuracy :
Training Accuracy :
Validation Accuracy :
|
|
Best estimate of validation accuracy for a generalised solution.
|
|
|
Comments about this data set what did you learn about it if anything
|
|