One of the most famous datasets is the MNIST dataset, which consists of handwritten digits, as shown in the following image. The dataset consists of 60,000 training and 10,000 test images:
The dataset is commonly used in image recognition problems to benchmark algorithms. The worst recorded error rate is 12%, with no preprocessing and using an SVM in a one-layer neural network. Currently, as of 2016, the lowest error rate is only 0.21%, using the DropConnect neural network, followed by a deep convolutional network at 0.23%, and a deep feedforward network at 0.35%.
Now, let's look at how to load the dataset.
