Learning Numbers With Neural Networks

• Using Jupyter Notebooks and the TensorFlow Library to create machine learning models for classifying handwritten digital images.
• Uses MNIST dataset of handwritten digits.

3 Sets of 5 experiments

1. First Set using raw (unchanged) data.
2. Second Set using scaled data.
3. Third Set using normalized data.

Each of the 5 experiments had the following characteristics (Control variables):

• Categorical-Cross-Entropy loss function.
• Adam optmizer.
• 10 iterations (epochs).
• 4 dense layered machine learning model.
• Layer 1 (input layer) of 784 nodes.

For Each of the 5 experiments, these were the independent variables:

• Experiment 1:
  • Layer 2: 4 nodes, linear activation function.
  • Layer 3: 4 nodes, linear activation function.
  • Layer 4: 10 nodes, linear activation function.
• Experiment 2:
  • Layer 2: 4 nodes, relu activation function.
  • Layer 3: 4 nodes, relu activation function.
  • Layer 4: 10 nodes, softmax activation function.
• Experiment 3:
  • Layer 2: 6 nodes, relu activation function.
  • Layer 3: 6 nodes, relu activation function.
  • Layer 4: 10 nodes, softmax activation function.
• Experiment 4:
  • Layer 2: 10 nodes, relu activation function.
  • Layer 3: 10 nodes, relu activation function.
  • Layer 4: 10 nodes, softmax activation function.
• Experiment 5:
  • Layer 2: 100 nodes, relu activation function.
  • Layer 3: 100 nodes, relu activation function.
  • Layer 4: 10 nodes, softmax activation function.

Conclusion

• Normalized data with relu and softmax activation functions gave most accurate results.
• Increasing the number of nodes at each layer improves accuracy only slightly.