A Simple AutoEncoder for images in the MNIST DATASET

The learning capability of the simplest autoencoder is explored in this work. We see if the self-supervised autoencoder can accurately learn to reconstruct the MNIST digits.

Useful parameters to consider tuning in this simple autoencoder

• Model Architecture: In this case we pick Multi-Layer Perceptrons of size [784, 128, 32 (latent representation size)] for the encoder and symmetrically for decoder.
• Learning rate scheduler : REDUCELRONPLATEAU with default factor and patience;
• Optimizer : Adaptive momentum with default betas and LR of 0.001
• Codeword represenation size
• misc: epoch, regulariztion etc.

Reconstruction results

Original

Reconstructed