sandbox for work and study
Convolutional Autoencoder, Search for similar images. main idea from: https://www.cs.toronto.edu/~hinton/science.pdf
Expirement with normal noise.
- input: image
- output: two vectors, first is class, second is noise. Each class has its own normal distribution with its mean and sigma. The experiment showed that in this way a very good separation of classes can be achieved. In the figure full_image_xception_noise_MAE.jpg you can see the cluster for each class from the test sample.
Fully Convolutional Network (FCN), main idea: https://arxiv.org/pdf/1605.06211.pdf