LeNet-5 is a classical neural network architecture that was successfully used on a handwritten digit recognition problem back in 1998. In principle, LeNet-5 was the first architecture that introduced the idea of applying several convolution layers before connecting to a fully-connected hidden layer. Before that, people would construct the features manually and then connect to a simple neural network with many hidden layers and neurons.

Here's the LeNet 5 architecture:

According to the paper, http://users.cecs.anu.edu.au/~Tom.Gedeon/conf/ABCs2018/paper/ABCs2018_paper_57.pdf, this model was able to achieve 99.05% accuracy, which is quite impressive considering the processing power available at the time wasn't good, and it has approximately 60,000 parameters. The architecture will look very similar to what we have worked with previously in this book. We'll use it to build a Java application for a handwritten digit recognition problem.