see https://arxiv.org/abs/1906.00446
see https://github.com/deepmind/sonnet/blob/master/sonnet/examples/vqvae_example.ipynb
I use Vector Quantizer to recreate 1-minute distributions of trading data. This distribution shows the volatility and positive/negative movement within a minute. If financial data can be simulated and datasets broadened, it could help to improve accuracy on existing regression or even reinforcement learning models.
Here, the model encodes an original into a latent space. Then the model decodes the latent space back to a distribution. Per the article, to create new distributions, another model must be made to find the relationships between latent space values. Those latent space value would then feed into the decoder.
