The standard example for machine learning these days is the MNIST data set, a collection of 70,000 handwriting samples of the numbers 0-9. now we predict which number each handwritten image represents.Each image is 28x28 grayscale pixels, so we treat each image as just a 1D array, or tensor, of 784 numbers.MNIST provides 60,000 samples in a training data set, 10,000 samples in a test data set, and 5,000 samples in a "validation" data set. We haven't talked about validation sets before, but their intent is to be used for model selection. So you'd use validation data to select your model, train the model with the training set, and then evaluate the model using the test data set.The training data, after we "flatten" it to one dimension using the reshape function, is therefore a tensor of shape [60,000, 784] - 60,000 instances of 784 numbers that represent each image. we define our architecture by 1 hidden layer and we use relu for activating nodes and we use 20 epochs and keep batch size of 100.