Handcrafted Features for DL Classification

In this project, we propose a new method to generate new realistic instances of a given dataset using a generative network only.
In our model, as in a standard architecture, we use linear layers and non-linear layers. However, we choose to fix the linear parameters and learn the non-linear parameters.
In this report, we show the algorithm results both on grayscale images (such as MNIST) and color images (such as CIFAR10).
The results that we obtained indicate a large potential for the architecture that we suggested, both for generative models and for variety of learning applications.
Furthermore, the results confirm that we can generate realistic images without using a discriminative network, what leads to a more robust and effortless training process.

Handcrafted Features for DL Classification
Handcrafted Features for DL Classification