MNIST Handwriting Recognition with TensorFlow

This project implements 3-layer linear neural network to classify the MNIST dataset using the TensorFlow. I aimed for more than 98% test set accuracy without overfitting.

Requirements

• Python 3.6
• TensorFlow
• matplotlib
• numpy

Results

If you read the MNIST successfully using tensorflow.examples.tutorials.mnist the output looks similar to below:

Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz

Number of training examples: 55000
Number of test examples: 10000

While repeating the mini-batch loop, the cost is displayed as below:

Cost after epoch 0: 0.476846
Cost after epoch 10: 0.033396
Cost after epoch 20: 0.011072
Cost after epoch 30: 0.005313
Cost after epoch 40: 0.004301
Cost after epoch 50: 0.002868
Cost after epoch 60: 0.003237
Cost after epoch 70: 0.002007
Cost after epoch 80: 0.001794
Cost after epoch 90: 0.002296

Cost is displayed as a graph with train and test accuracy after gradient descent with the Adam Optimizer is finished.

Train Accuracy: 0.999982
Test Accuracy: 0.9842