This document is about my ideas what we can do, based on what I've read in papers or anywhere on the internet.

## Augmentation

### Use GAN
Gan generates new samples of data. Generator is trained alongside descriminator. After that we have a generator capable of generating new data. I mean it is not augmentation of dataset, it is creating whole new dataset.
Tho the generator needs some input. And that is the question, what should it be. 
* One option is to use labels with some random noise as input. 
* This paper shows different approach [[Augmentation methods#Generative Adversarial Networks-Based Data Augmentation for Brain–Computer Interface(2020)]]. Based on that I propose method where we would extract something like **movement-specific feature vector** which would together with noise be input to generator.

### Contrastive learning
youtube video with explanation: https://www.youtube.com/watch?v=UqJauYELn6c

## Feature Extraction
I think we should try variational autoendcoders, and some novel architecture, like VQVAE(vector quantization should bring narrower space for classifier at the end).