Handwritten Digit Generator
Working off the MNIST database, I used TensorFlow to create a conditional variational auto-encoder that was able to generate its own images of handwritten digits. After sampling a latent vector from a random normal distribution, the network's generated images had an accuracy rating of 98%, when judged by Brown's MNIST discriminator. None of the images on the right were given to the program, but generated by the computer after it studied over 60,000 samples.
These images are created by passing input to a neural network (the encoder), which generates a compressed latent vector that is then passed through another neural network (the decoder) to generate a new image. Once trained off of a dataset of example images, the networks are able to generate realistic output by sampling latent space.