There's more...

Our one hidden layer multilayered perceptron gave an accuracy of 84.45 on training data and 92.1 on test data. It is good but not good enough. The MNIST database is used as a benchmark for classification problems in machine learning. Next, we see how using TensorFlow's built-in optimizer affects network performance.