Quite accidentally I found the picture below showing my research work in 1994 when I trained a neural network to identify the land use types from a satellite image. The neural network had a structure of 3 input neurons, corresponding to the blue, green, and red bands of a Landsat TM image, 14 hidden neurons, and 7 output neurons, representing seven land use types including the cornfields, wheat fields, water bodies, impervious surfaces, etc. The number of training data points is 65.
I coded everything from scratch including the algorithm and the user interface using C++. There were no prebuilt packages to use, no internet to search the information, and no such thing as asking questions on social media. The computer used had a 486 CPU, a 32-megabyte memory, and a hard drive with less than 500-megabyte storage.
How things have changed in 27 years! With a deep learning neural network, such as a convolutional neural network, many layers of neurons serving different purposes can be stacked together to form a complex structure and collectively perform recognitive tasks that were unthinkable in the past. And all these can be done with a few lines of Python scripts and the execution is often done in a cloud computing environment with virtually unlimited computation and storage resources. The progress is astonishing.