Dataset - https://www.kaggle.com/c/digit-recognizer/data

Digit-Recognizer

Here we focus on identifying handwriting images via CNN. The CNN architecture used is -

[[Conv2D->relu]*2 -> MaxPool2D -> Dropout]*2 -> Flatten -> Dense -> Dropout -> Out

Conv2D is the first layer is which is used as a filter. The filters here can be seen as a transformation of the image. The CNN then maps these features from the transformed images.

Next layer is the MaxPool2D layer which is used to reduce computational cost and reduce overfitting. It is more like a downsampling filter which picks it’s 2 neighbouring pixels and chooses the maximum value. We pick the pooling size and as the size increases the downsampling becomes more significant.

Dropout is a regularization method, where a proportion of nodes in the layer are randomly ignored by setting their weights to zero for each training sample. This drops randomly a proportion of the network and forces the network to learn features in a distributed way. This technique also improves generalization and reduces the overfitting.

We have used ReLu as the activation function. 'relu' is the rectifier (activation function max(0,x). which is used to add nonlinearity to the network.

The Flatten layer is used to convert the final feature maps into a one single 1D vector. This flattening step is needed so that you can make use of fully connected layers after some convolutional/maxpool layers. It combines all the found local features of the previous convolutional layers.

We have considered epochs to be 2 to computational time, it can be increased to 30 for a higher accuracy.