II. Neural Networks
Idea: Use a three-layer perceptron network, where each layer has its own weights and bias functions, to classify data into six different postures.
Steps:
Initialize weights and bias units to small values.
Input layer uses a logsin transfer function with 56 inputs.
Hidden layer has 45 hidden units (uses a logsin transfer function) and output layer (purelin transfer function) has 6 output units.
Uses gradient descent backpropagation to find weights and bias units of each layer.
Simulate network for a test set.
.Results.
Neural Network Params | |||||||
Hidden Units | 24 | 24 | 40 | 60 | 50 | 40 | 45 |
Transfer Functions | logsig,logsig,logsig | logsig, logsig, purelin | logsig, logsig, purelin | logsig, logsig, purelin | logsig, logsig, purelin | logsig, logsig, purelin | logsig, logsig, purelin |
Epochs | 500 | 500 | 500 | 500 | 500 | 500 | 500 |
Learning Rate | 0.1 | 0.001 | 0.0001 | 0.0001 | 0.00001 | 0.000001 | 0.0000001 |
% Recognition Train | 41.50% | 50.67% | 61.83% | 60.17% | 60.00% | 60.00% | 61.17% |
% Recognition Test | 35.33% | 39.67% | 48.00% | 47.00% | 46.33% | 47.33% | 54.67% |
Neural Network Params | |||||||
Hidden Units | 45 | 47 | 45 | 45 | 40 | 55 | 65 |
Transfer Functions | logsig, logsig, purelin | logsig, logsig, purelin | logsig, logsig, purelin | tansig, logsig, purelin | tansig, logsig, purelin | tansig, logsig, purelin | tansig, logsig, purelin |
Epochs | 500 | 500 | 500 | 700 | 700 | 700 | 700 |
Learning Rate | 0.000000001 | 0.0000001 | 0.0000001 | 0.0000001 | 0.0000001 | 0.0000001 | 0.0000001 |
% Recognition Train | 59.50% | 55.50% | 60.50% | 61.33% | 61.17% | 63.67% | 60.67% |
% Recognition Test | 46.67% | 46.67% | 52.00% | 50.67% | 46.33% | 47.00% | 53.00% |
Neural Network Params | ||
Hidden Units | 75 | 100 |
Transfer Functions | tansig, logsig, purelin | tansig, logsig, purelin |
Epochs | 700 | 700 |
Learning Rate | 0.0000001 | 0.0000001 |
% Recognition Train | 61.83% | 60.83% |
% Recognition Test | 50.67% | 46.00% |