Latent Space Auto Encoder

Hello,

I am wondering about the autoencoder, especially the embedder that transforms the data into the latent space. If I understood it correctly, the embedders task is to compress the dimensionality of the original data into a lower dimension in order to learn the underlying structure.

But lets say we have the stock data (6 columns), a rolling windows sice of 24 and a batch sice of 50. We give the embedder data of the shape (50,24,6). In the code we use 24 neurons per layer... so that the output of the embedder is 50,24,24....

Wouldn't that imply that instead of compressing the data to a lower dimensional space, we kind of extend the latent space? Dont we need a bottleneck layer in th autoencoder of less than 6 in order to achieve the dimensionaliy reduction we are looking for?

Thanks in advance for your response. Kind regards, Franz

ydataai / ydata-synthetic

Latent Space Auto Encoder #165