DigitRecognition

Update

User interface still in development!!!

Description

A fully connected neural network recognizing hand-written digits with NumPy.

Use MNIST .csv dataset to train. (Ignored in the repository)

Use Pygame to visualize the process and implement the drawing pad interface.

Use Threading to separate model calculation and screen update.

Use Tkinter to load and save the trained model.

Use Pillow to process image.

Installation

• NumPy: pip install numpy

• Pygame: pip install pygame

• Tkinter: pip install tk

• Pillow: pip install pillow

• Update the MNIST_path in NeuralNetwork.py to the corresponding training dataset .csv file.

  MNIST_path = './MNIST/mnist_test.csv'

Usage

Run NeuralNetwork.py in the terminal.

python .\NeuralNetwork.py

Code

• NeuralNetwork.py

  Main code file.

  Pygame running code.

• Layer.py

  Class Hidden_Layer, Output_Layer

  Include .forward(), .backward(), .learn() method for forward propagation, backward propagation, adjusting weight & bias.

• ActivationFunction.py

  Activation functions: $ReLU$, $Sigmoid$, $\tanh$, $Softmax$.

  Loss function: $Cross-entropy$

  and their derivatives.

• PygameClass.py

  Class PAINT: Drawing canvas Class TEXT: Text box Class BUTTON: Clickable button

Math

Math equations such as matrix cannot properly display on GitHub.

This README file is written in VS Code.

Please use VS Code or other Markdown reader to view.

• Prior Knowledge

  Vector & Matrix:

  • Matrix Multiplication
  • Transpose

  Multivariable Calculus:

  • Partial Derivative
  • Gradient

  One-Hot Encoding

• Variable

  $w$: weight

  $b$: bias

  $a$: activation

  $z$: unormalized activation (weighted sum)

  $L$: output layer

  $y$: desired output

  $l$: loss

• Activation Function

  Rectified linear unit:

  $ReLU(x) = \left{ \begin{matrix} x & x>0 \ 0 & x\leqslant 0 \end{matrix} \right.$

  Derivative of ReLU:

  $ReLU'(x) =\left{ \begin{matrix} 1 & x>0 \ 0 & x<0 \end{matrix} \right.$
  
  

  Sigmoid:

  $\sigma(x) = \frac{1}{1+e^{-x}}$

  Derivative of sigmoid:

  $\sigma'(x) = \sigma(x)(1-\sigma(x)) = \frac{1}{1+e^{-x}} (1-\frac{1}{1+e^{-x}})$
  
  

  $\tanh$:

  $\tanh(x) = \frac{e^x-e^{-x}} {e^x+e^{-x}}$

  Derivative of $\tanh$

  $\tanh'(x) = 1-\tanh^2(x)$
  
  

  Softmax:

  $softmax(x)i = \frac{e^{x_i}} {\sum^K{j=1} e^{x_j}}$

  Derivative of softmax:

  Jacobian matrix(To be updated...)

• Loss Function

  Cross-entropy:

  $H(p,q)=-\sum p(x)\log q(x)$
  
  

  Mean squared error:

  $MSE=\frac{1} {n} \sum^n_{i=1} (y_i-\hat{y}_i)^2$

  $y$: desired value

  $\hat{y}$: predicted value

• Symbol Notation

  $w\cdot b$: dot product / matrix multiplication

  $w \circ b$: Hadamard product / element-wise product

  $k \times j$: matrix / vector dimension, $k$ rows, $j$ columns

  $a^T$: transpose

• Neural Network

  Forward Propagation

  Layer Input

  $a^I_{1 \times I} = \begin{bmatrix} a^I_1 & a^I_2 & \cdots & a^I_I \end{bmatrix}$

  Layer 1

  $w^1_{I \times m} = \begin{bmatrix} w^1_{1,1} & w^1_{1,2} & \cdots & w^1_{1,m} \ w^1_{2,1} & w^1_{2,2} & \cdots & w^1_{2,m} \ \vdots & \vdots & \ddots & \vdots \ w^1_{I,1} & w^1_{I,2} & \cdots & w^1_{I,m} \end{bmatrix}$

  $b^1_{1 \times m} = \begin{bmatrix} b^1_1 & b^1_2 & \cdots & b^1_m \end{bmatrix}$

  $z^1_{1 \times m} = a^I_{1 \times I} \cdot w^1_{I \times m} + b^1_{1 \times m}$

  $a^1_{1 \times m} = ReLU(z^1_{1 \times m})$

  Layer 2

  $w^2_{m \times k} = \begin{bmatrix} w^2_{1,1} & w^2_{1,2} & \cdots & w^2_{1,k} \ w^2_{2,1} & w^2_{2,2} & \cdots & w^2_{2,k} \ \vdots & \vdots & \ddots & \vdots \ w^2_{m,1} & w^2_{m,2} & \cdots & w^2_{m,k} \end{bmatrix}$

  $b^2_{1 \times k} = \begin{bmatrix} b^2_1 & b^2_2 & \cdots & b^2_k \end{bmatrix}$

  $z^2_{1 \times k} = a^1_{1 \times m} \cdot w^2_{m \times k} + b^2_{1 \times k}$

  $a^2_{1 \times k} = ReLU(z^2_{1 \times k})$

  Layer Output

  $w^O_{k \times O} = \begin{bmatrix} w^O_{1,1} & w^O_{1,2} & \cdots & w^O_{1,O} \ w^O_{2,1} & w^O_{2,2} & \cdots & w^O_{2,O} \ \vdots & \vdots & \ddots & \vdots \ w^O_{k,1} & w^O_{k,2} & \cdots & w^O_{k,O} \end{bmatrix}$

  $b^O_{1 \times O} = \begin{bmatrix} b^O_1 & b^O_2 & \cdots & b^O_O \end{bmatrix}$

  $z^O_{1 \times O} = a^2_{1 \times k} \cdot w^O_{k \times O} + b^O_{1 \times O}$

  $a^O_{1 \times O} = softmax(z^O_{1 \times O})$

  Loss

  $y = \begin{bmatrix} y_1 & y_2 & \cdots & y_O \end{bmatrix}$ (One-Hot Encoding)

  $l = -\sum^O_{j=1} y_j\ln a^O_j = -y_1ln a^O_1 - y_2ln a^O_2 - \cdots - y_Oln a^O_O$ (Cross-entropy)

  Backward Propagation

  Layer Output

  $\begin{align} \notag {\frac{\partial l} {\partial a^O}}_{1 \times O} & = & \begin{bmatrix} \frac{\partial l} {\partial a^O_1} & \frac{\partial l} {\partial a^O_2} & \cdots & \frac{\partial l} {\partial a^O_O} \end{bmatrix} \ \notag & = & \begin{bmatrix} -\frac{y_1} {a^O_1} & -\frac{y_2} {a^O_2} & \cdots & -\frac{y_O} {a^O_O} \end{bmatrix} \end{align}$

  This is a Jacobian matrix:

  $\begin{align} \notag {\frac{\partial a^O} {\partial z^O}}_{O \times O} & = & \begin{bmatrix} \frac{\partial a^O_1} {\partial z^O_1} & \frac{\partial a^O_1} {\partial z^O_2} & \cdots & \frac{\partial a^O_1} {\partial z^O_O} \ \frac{\partial a^O_2} {\partial z^O_1} & \frac{\partial a^O_2} {\partial z^O_2} & \cdots & \frac{\partial a^O_2} {\partial z^O_O} \ \vdots & \vdots & \ddots & \vdots \ \frac{\partial a^O_O} {\partial z^O_1} & \frac{\partial a^O_O} {\partial z^O_2} & \cdots & \frac{\partial a^O_O} {\partial z^O_O} \end{bmatrix} \ \notag & = & \begin{bmatrix} a^O_1(1-a^O_1) & -a^O_1a^O_2 & \cdots & -a^O_1a^O_O \ -a^O_1a^O_2 & a^O_2(1-a^O_2) & \cdots & -a^O_2a^O_O \ \vdots & \vdots & \ddots & \vdots \ -a^O_1a^O_O & -a^O_2a^O_O & \cdots & a^O_O(1-a^O_O) \end{bmatrix} \end{align}$

  $\because y$ is a One-Hot, $\sum^O_{j=1}y_j=1$

  $\therefore$

  $\begin{align} \notag {\frac{\partial l} {\partial z^O}}{1 \times O} & = & {\frac{\partial l} {\partial a^O}}{1 \times O} \cdot {\frac{\partial a^O} {\partial z^O}}{O \times O} \ \notag & = & \begin{bmatrix} -y_1+a^O_1\sum^O{j=1}y_j & -y_2+a^O_2\sum^O_{j=1}y_j & \cdots -y_O+a^O_1\sum^O_{j=1}y_j \end{bmatrix} \ \notag & = & \begin{bmatrix} a^O_1-y_1 & a^O_2-y_2 & \cdots a^O_O-y_O\end{bmatrix} \ \notag & = & a^O-y \end{align}$

  $\begin{align} \notag {\frac{\partial l} {\partial w^O}}{k \times O} & = & {\frac{\partial z^O} {\partial w^O}}{k \times 1} \cdot {\frac{\partial l^O} {\partial z^O}}_{O \times O} \ \notag & = & a^{2T} \cdot \frac{\partial l} {\partial z^O} \end{align}$

  $\because \frac{\partial z^O} {\partial b^O}$ is a Jacobian matrix and is a identity matrix

  $\therefore$

  $\begin{align} \notag {\frac{\partial l} {\partial b^O}}{1 \times O} & = & {\frac{\partial l} {\partial z^O}}{1 \times O} \cdot {\frac{\partial z^O} {\partial b^O}}_{O \times O} \ \notag & = & \frac{\partial l} {\partial z^O} \cdot 1 \end{align}$

  Layer 2

  $\begin{align} \notag {\frac{\partial l} {\partial a^2}}{1 \times k} & = & {\frac{\partial l} {\partial z^O}}{1 \times O} \cdot {\frac{\partial z^O} {\partial a^2}}_{O \times k} \ \notag & = & \frac{\partial l} {\partial z^O} \cdot w^{OT} \end{align}$

  $\begin{align} \notag {\frac{\partial l} {\partial z^2}}{1 \times k} & = & {\frac{\partial l} {\partial a^2}}{1 \times k} \cdot {\frac{\partial a^2} {\partial z^2}}_{k \times k} \ \notag & = & \frac{\partial l} {\partial a^2} \circ ReLU'(z^2) \end{align}$

  $\begin{align} \notag {\frac{\partial l} {\partial w^2}}{m \times k} & = & {\frac{\partial z^2} {\partial w^2}}{m \times 1} \cdot {\frac{\partial l} {\partial z^2}}_{1 \times k} \ \notag & = & a^{1T} \cdot {\frac{\partial l} {\partial z^2}} \end{align}$

  $\because \frac{\partial z^2} {\partial b^2}$ is a Jacobian matrix and is a identity matrix

  $\therefore$

  $\begin{align} \notag {\frac{\partial l} {\partial b^2}}{1 \times k} & = & {\frac{\partial l} {\partial z^2}}{1 \times k} \cdot {\frac{\partial z^2} {\partial b^2}}_{k \times k} \ \notag & = & \frac{\partial l} {\partial z^2} \cdot 1 \end{align}$

  Layer 1

  $\begin{align} \notag {\frac{\partial l} {\partial a^1}}{1 \times m} & = & {\frac{\partial l} {\partial z^2}}{1 \times k} \cdot {\frac{\partial z^2} {\partial a^1}}_{k \times m} \ \notag & = & \frac{\partial l} {\partial z^2} \cdot w^{2T} \end{align}$

  $\begin{align} \notag {\frac{\partial l} {\partial z^1}}{1 \times m} & = & {\frac{\partial l} {\partial a^1}}{1 \times m} \cdot {\frac{\partial a^1} {\partial z^1}}_{m \times m} \ \notag & = & \frac{\partial l} {\partial a^1} \circ ReLU'(z^1) \end{align}$

  $\begin{align} \notag {\frac{\partial l} {\partial w^1}}{I \times m} & = & {\frac{\partial z^1} {\partial w^1}}{I \times 1} \cdot {\frac{\partial l} {\partial z^1}}_{1 \times m} \ \notag & = & a^{IT} \cdot {\frac{\partial l} {\partial z^1}} \end{align}$

  $\because \frac{\partial z^1} {\partial b^1}$ is a Jacobian matrix and is a identity matrix

  $\therefore$

  $\begin{align} \notag {\frac{\partial l} {\partial b^1}}{1 \times m} & = & {\frac{\partial l} {\partial z^1}}{1 \times m} \cdot {\frac{\partial z^1} {\partial b^1}}_{m \times m} \ \notag & = & \frac{\partial l} {\partial z^1} \cdot 1 \end{align}$