End-to-end Sinkhorn Autoencoder with Noise Generator
- Deja, Kamil et al
- arXiv:2006.06704
 | : |
 | : : Samples of generated numbers from the model trained on MNIST dataset(left) and interpolation between different classes(right). Our conditional end-to-end sinkhorn autoencoder creates various images. During interpolation it does not create examples from outside of original classes even for invalid conditional parameters. |
 | figure |
 | figure |
 | Schematic visualisation of end-to-end sinkhorn autoencoder processing (left). TSNE visualisation of latent space for mnist dataset (right). Our conditional e2e sinkhorn autoencoder (top) and conditional VAE (bottom). Our model does not restrict latent space to the normal distribution, therefore classes may be even linearly separable. |
 | Schematic visualisation of end-to-end sinkhorn autoencoder processing (left). TSNE visualisation of latent space for mnist dataset (right). Our conditional e2e sinkhorn autoencoder (top) and conditional VAE (bottom). Our model does not restrict latent space to the normal distribution, therefore classes may be even linearly separable. |
 | Schematic visualisation of end-to-end sinkhorn autoencoder processing (left). TSNE visualisation of latent space for mnist dataset (right). Our conditional e2e sinkhorn autoencoder (top) and conditional VAE (bottom). Our model does not restrict latent space to the normal distribution, therefore classes may be even linearly separable. |
 | Architecture of the sinkhorn autoencoder with neural network as an explicit noise generator. Red arrows indicate the gradient flow. Reconstruction Loss $L_2$ is backpropagated through decoder and encoder, while sinkhorn loss $L_1$ is propagated in two directions to encoder and noise generator. Encoder network is optimised with a sum of $L_1$ and $L_2$ losses. |
 | Examples of calorimeters response simulations with different methods. Although results from GAN are visually sound with collisions, model was not able to properly capture relations from conditional values. Our solution did not reproduce all of residual values, it outperformed other methods in terms of accuracy for the most significant centre of collision. |
 | : |
 | : |
 | : TSNE visualisation of generated examples (blue) and original mnist data (red), processed through LeNet network. DCGAN reproduces the whole data distribution well, while VAE additionally produces images from outside of real data distribution. Our solution (right) generates only examples within true data distribution but does not properly reproduce the whole variety. : Caption not extracted |
 | figure : Results comparison on CelebA dataset. For competitive solutions we include the best of reported result.figure |
 | : |
 | : : Samples of generated numbers from the model trained on MNIST dataset(left) and interpolation between different classes(right). Our conditional end-to-end sinkhorn autoencoder creates various images. During interpolation it does not create examples from outside of original classes even for invalid conditional parameters. |
 | figure |