<<<<<<< HEAD TODO :

The model architecture here is dead simple. I might change it in favour of a simple CNN if it doesn't perform but the architecture is this:

[784] Input MNIST vector --> [196] Layer 1 --> [49] Hidden Layer --> [10] Output Layer

Layer 1

• Input Dimensions: [784]
• Output Dimensions: [196]
• Weight Dimensions: [784,196]
• Bias Dimensions: [196]
• Activation Function: Softmax

Hidden Layer

• Input Dimensions: [196]
• Output Dimensions: [49]
• Weight Dimensions: [196,49]
• Bias Dimensions: [49]
• Activation Function: Softmax

Output Layer

• Input Dimensions: [49]
• Output Dimensions: [10]
• Weight Dimensions: [49,10]
• Bias Dimensions: [10]
• Activation Function: Softmax

The cost function used is the cross entropy cost function. Apparently it works well for NN's that use softmax activation functions (which this one does).

This is a very basic neural network built in Rust purely to understand rust better. There is no abstraction/deep learning library/fancy kernel that maps a classifier into a higher dimensional space. This code is probably ugly, badly written and slower than it code be.

The model architecture here is dead simple.

[28,28] Input MNIST vector --> [784,1] Input Layer --> [56,1] Hidden Layer --> [10,1] Output Layer

The cost function used is the cross entropy cost function. Apparently it works well for NN's that use sigmoids activation functions (which this one does).

Work in progress

6fcefd962036c50a376caef616a6e97d5b9657ab