Auto Encoders are a special type of Neural neworks architecture that can be utilized for various applications like simple compression of data(basic auto-encoder), de-noising of data(de-noising auto-encoder) and also generating new samples (Generative AI) of data similar to our dataset(Variational auto-encoders VAEs).

Here, I used the Fashion MNIST and Handwritten-digits MNIST data to implement the above mentioned applications of Auto-Encoders.

Compression and Reconstruction

Noise is intentionally added to the data but the model is trained to get the original de-noised image.

Generate new samples from the end-to-end model by just getting a sample from the latent-space of the trained VAE.

• Generative results from VAE after 10th epoch

• The latent-space of the trained VAE

We can see a nice continuous space representing the distribution from one number to another.