Autoencoder to compress and reconstruct handwritten digits from the MNIST dataset. Note, the model starts with pre trained weights for the purpose of this demo.
This demo showcases a pre-trained autoencoder neural network that compresses MNIST digit images down to a compact latent representation of just 49 dimensions, before reconstructing them. Here's the process:
This technique is useful for dimensionality reduction, feature learning, and potentially generating new digit images.
The demo will cycle through different test images.
Models trained with different activation functions are available to compare. Each model was trained for 20,000 steps with the same hyperparameters.
Original Image | Reconstructed Image |
---|---|
The autoencoder consists of an encoder network that compresses the input, and a decoder network that reconstructs it.