Artifical Neural Network From Scratch in Mnist-Digit Dataset.

ANN Classifier used to classify MNIST Digit.

Concepts/Things implemented in this Notebook:

1. Built Mnisit Digit Classifier with three different layers. (one with 1 layer, 2 layer and 4 layer).
2. Trained with different optimizers like GD with Momentum, AdaGrad and Adam.
3. Trained with GPU to boost training.
   P.S.: Everything here is built with numpy, python and pandas."Scratch".

Findings/Results:

1. Notebook-1 :

Layers	Optimizer	Epoch	Learning Rate	Training Accuracy	Testing Accuracy
1	Gradient Descent	10	0.01	74.56%	73.899%
2	Gradient Descent	20	0.01	89.799%	88.9111111%
4	Gradient Descent	50	0.01	74.56%	73.899%

Note: Since, all the training is done in cpu. So, while training, go get yourself a cup of coffee.☕😃. Later we'll try in GPU.😉(Present in the notebook 3)

2. Notebook-2:

Layers	Optimizer	Batch Size	Epoch	Learning Rate	Momentum Parmeter	Training Accuracy	Testing Accuracy
2	Mini-Batch Gradient Descent with Momentum	64	50	0.01	0.9	85.232%	84.649%

3. Notebook-3:

Layers	Optimizer	Batch Size	Epoch	Learning Rate	Epsilon	Training Accuracy	Testing Accuracy
4	Mini-Batch Gradient Descent with AdaGrad	200	30	0.01	10e-8	81.085%	__%

Three ANN of different layers is used to classify images.

1. One Hidden Layer

---

2. Two Hidden Layer

---

3. Four Hidden Layer

---

---

Note : In this notebook, i've trained the model only for few epochs. Since i'm using cpu to train the model, it is going to take time(a lot in more layers ANN). So, if you want to get boost in the accuracy, train the model in more epochs.

We'll use CNN in this same MNIST DIGIT dataset.
You can find repo Here.