- To put in simple terms, Autoencoders are simply feature extractors with additional task of
reconstructiong the given input which maybe images or videos with minimum losses. - In Modern neural networks the autoencoders are utilized for many other purposes like that of
image segmentation (refer U-nets), image denoising, image completion or also in modern reinforcement learning algorithms (Graph neural networks).
- Convolutional Autoencoders work well untill scaled down to 1x1 feature levels.
- One can not be hopeful of generating a new image using only the trained conv-generator.
- Making the autoencoder conditional does not help to improve the quality of images reproduced.
This is my best attempt at reproducing the images using a conv-autoencoder for code size of 2x2 features.
| Dataset | STL10 |
|---|---|
| Loss Criterion | MSE Loss(0.0028) |
| Network Type | Simple Convolutional Network |
- Variational Autoencoders are a link between the autoencoders and generative networks. Simply, they are
also able to sample and create new images better than conv-autoencoders. - It has a additional KL-Divergence loss which aims to remove inter-dependibility of each feature of the bottleneck part and spread it in a lower bound known to resample it.
- Here I wanted to use only 40 features so Image reconstructions are not very clear but should be better with
more features on last part of encoder - I have also made only the decoder part of network conditional by supplying additional labels. (previous experiments
of making the whole network conditional also gives same results). - Tried using Resnet Type network for encoder but should work fine with normal encoders also.
| Dataset | Cifar10 |
|---|---|
| Loss Criterion | Kl+MSE loss (~13000) |
| Feature size | 40 |
| Network Type | VAE |
I know it is not the best but it created this only using 40 features and also for a complex dataset then MNIST.
- Image completion.
- Denoising images.