In this talk I demonstrate how the rate-distortion theory works for data compression using autoencoder. The encoder and decoder are both convolutional neural network (CNN) architectures and the demo dataset is from Stanford's Cars Dataset.