Learning about auto encoders and latent spaces
Experiments TODO:
Standard autoencoder on MNIST:
- Experiment with the trivial autoencoder for the identity mapping case (photometric loss, L = ||x - net(x)||^2)
- Sparse autoencoders:
- Identity function:
- Noisy inputs to sparse autoencoder (robustness to noise, resulting in a richer latent space)
- Image denoising
- Losses to try:
- L1 regularized
- KL divergence regularization
Perhaps:
- Visulization of activations
- VAE on MNIST (if I have time)
- Comparison of PCA vs Overcomplete autoencoder (compare latent space vs PCA)
(TODO) List of resources used partitioned by category. For later reference.