Скачать или смотреть Bad Apple!! but it's learned by a Vector Quantized Variational Autoencoder for reconstruction.

A Vector Quantized Variational Autoencoder (VQ-VAE) is an application of machine learning which leverages a mechanism called vector quantization to map continuous latent representations into discrete embeddings.

The purpose of the VQ-VAE is to learn a meaningful compressed representation of our original data. There are two models the Encoder and Decoder. The encoder takes the original video as input and outputs a compressed latent representation, the decoder takes the compressed latent representation and tries to reconstruct the original.

We go from Data to Encoder to Bottleneck (Compressed Latent Space) to Decoder to Reconstruction of our data.

In this case, we take the original video as input, pass it through our VQ-VAE which then compresses it into a latent space before outputting a reconstruction.

The compression ratio is pretty high in this case! It's trained like this:

Model
Save frame/audio
Calculate Loss
Loss.backward()

So the video is validation at different stages in training :D

#machinelearning #artificialintelligence #generativeai #ai #aiart #ml #badapple #touhou #touhouproject #remix