https://github.com/ayooshkathuria/mnist-handwritten-digit-recognition

Classification of MNIST Handwritten Digits Database using Deep Learning
https://github.com/ayooshkathuria/mnist-handwritten-digit-recognition

Last synced: about 2 months ago
JSON representation

Classification of MNIST Handwritten Digits Database using Deep Learning

• Host: GitHub
• URL: https://github.com/ayooshkathuria/mnist-handwritten-digit-recognition
• Owner: ayooshkathuria
• Created: 2016-11-11T21:44:28.000Z (about 8 years ago)
• Default Branch: master
• Last Pushed: 2017-04-10T01:06:43.000Z (almost 8 years ago)
• Last Synced: 2024-11-26T07:19:06.090Z (about 2 months ago)
• Language: Python
• Homepage:
• Size: 92.9 MB
• Stars: 7
• Watchers: 1
• Forks: 3
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
##Classification of MNIST Handwritten Digits Database using Deep Learning

This repository contains code meant to classifiy MNIST Handwritten Digits usind Neural Networks. I have used the data from 

http://deeplearning.net/data/mnist/ as well as an algorithmically expanded version of this datset for training the neural 

networks. There are 3 code files which are described below.

####NeuralNetBasic.py

A very simple neural network implemented in python using Stochastic Gradient Descent and Backpropogation. 

####NeuralNetOptimised

An optimised version of neural network above. Incorporates cross-entropy cost function instead of quadratic cost function

L2-regularisation, an algorithmically expanded dataset and better support for performance analysis. The weights have been 

initialised with mean zero and standard deviation 1/sqrt(Number or outputs) rather than 1.

####NeuralNetTheano

Unlike previous two versions, this neural network is implemented using Theano. It also incorporates a couple of convolutional 

layers and a softmax output layer in addition to fully connected layers. Dropout has also been implemented in fully connected 

layers to address the problem of overfitting.

The best efficiency I obtained was 98.87% with a couple of convolutional layers, a fully connected layer of 640 neurons and an 

output softmax layer of 10 neuron, learning rate of 0.1. The network trained over 50 epochs, and took a long while on my MacBook Air (I'm never doing that again).

This code hs been derived from code samples from the book http://neuralnetworksanddeeplearning.com/ authored by Michael Nielson. If you're looking for an introduction to deep learning, the book can be a great starting point.