Skip to content

Notes

Workflow Load dataset or sample data from EEG

Notes

We will use supervised learning to learn the EEG → Actions decoder. First, we will use publicly available and labeled datasets to train a NN.

Afterward, we try to train a NN using our own datasets, sampled with the cap.

The EEG will generate "noisy" action labels due to the noise in the EEG. The idea is to check if the CEILing framework can be used with these noisy labels.

EEG

Learning NN

We will first sample and store labeled datasets and have offline supervised learning of an NN.

There can be two sources of data: - Braindecode public available datasets - Our own sampled dataset using the EEG cap + gamepad

-[ ] Do we need to push the EEG data to ROS?

Using NN to decode actions

We want to get a "live stream" of data (EEG) sampling, "decode" and action, and then feed it to the Action-Motion policy from the CEILing.

Sampling from "Cap"

  • PC password: neuro
  • SW password: 0000
  • Sampling rate: 1 kHz
  • Amplifier: EEG64-CY-261
  • Raw data.

Action → Motion

The second step is to use the CEILing framework to train Actions to Motion. To perform the feedback we have:

  • Action → Motion
  • Evaluative feedback (human): ??
  • Corrective feedback (human): ??
  • Comparison feedback (human): ??

QUESTIONS

  • How to get the event_codes from the dataset?
  • Create custom dataset with my events?
  • I am creating an MNE dataset from my dataset, so then I can create an annotated BrainDecode dataset, but as far as I understand, I can only generate discrete events "either 1 or -1", not "0.9". Could this be a problem for the future, when we integrate the robot?