Simple notebooks to classify MNIST dataset using TensorFlow

There are two notebooks namely MNIST-CNN and MNIST-DNN in this repository. MNIST-DNN is a notebook which implements a Fully Connected Layer network to recognize the digits, giving a decent performance of around 97-98 %. In this I do the task in 2 ways : First using an explicitly defined generic function which generates minibatches and the Second using the next_batch feature of mnist data provided by TensorFlow

MNIST-CNN is a notebook which implements a Convolution Layer - FC Layer - Softmax architecture to recognize the digits. Note that this uses the relatively new tf.estimator API. Also the CNN gives a relatively bad performance of about 95-96 %, but this can be improved by using techniques like using better optimization algorithms like Adam instead of Stochastic Gradient Descent, running for more epochs, decreasing dropout probability, adding more layers, etc.