Introduction

MNIST ("Modified National Institute of Standards and Technology") is the de facto “Hello World” dataset of computer vision. Since its release in 1999, this classic dataset of handwritten images has served as the basis for benchmarking classification algorithms. As new machine learning techniques emerge, MNIST remains a reliable resource for researchers and learners alike.

Problem Statement

In this, we aim to correctly identify digits from a dataset of tens of thousands of handwritten images.

Hardware

1. 940MX GPU with CUDA 10.1
2. 8GB RAM at 2133 Hz.
3. i5 @2.4 Ghz with 4 cores.

Process

The process is divided into 3 parts.

1. Add Gaussian noise to dataset.
2. Create Autoencoder to denoise the image.
3. Create a Dense layer model to identify Digits.

Why Gaussian Noise

• It helps to prevent GAN attacks.
• It can be used for dimensionality reduction.

1. Adding Gaussian Noise

   1. How to add gaussian noise
   2. Here is the link to my the code. Link

2. Create Autoencoder

   1. Encoder - Input image => CONV => RELU => BN => CONV => RELU => BN => Flatten => Dense
   2. Decoder - Output Encoder image => CONV_TRANSPOSE => RELU => BN => CONV_TRANSPOSE => RELU => BN => CONV_TRANSPOSE => Sigmoid
   3. Autoencoder - Encoder + Decoder

3. Dense Layers model for prediction

   1. Dense => Dropout => Dense => Dropout => Dense => Dropout => Dense => Dense
   2. Output one hot encoded numbers between 0-9

Property	Value
Epochs(Autoencoder)	100
Epochs(Dense)	200
Optimizer(Autoencoder)	Adam
Optimizer(Dense)	Adam

Property	Training	Testing
No. of Images	60,000	10,000
Time (Autoencoder)	2 hrs	23 mins
Time (Digit Prediction)	20 mins	4 mins

Loss vs Epoch for Autoencoder

Loss vs Epoch for Predictor

Updated Model and Performance Metric:

AutoEncoder Sample

Autoencoder Loss

Dense Accuracy

Dense Loss