README.md

# LDS_BadElectrodes
The repository is for the ECE69500 - Generative Models Final Project.
The project is "Good vs Bad Electrodes in SEEG Data"

# SEEG Data:
This is neural data collected directly from the subject by inserting a long and thin electrode in the subject's brain. The electrode can be used to measure and record brain/neuron activity. This data can be used for tasks such as speech decoding directly from the neural activity.

# Bad Electrodes:
A problem arises when some of the electrodes are 'bad'. These electrodes don't measure data that can be useful for us, this is because the signal being outut by the electrode is either:
  1. Noisy
  2. Fluctuating
  3. Has repeating patterns

Good electrode
![image](https://media.github.itap.purdue.edu/user/6620/files/db5e9cc4-2537-4a76-93f9-1ff6a88450d3)

Bad electrode
![image](https://media.github.itap.purdue.edu/user/6620/files/7d55e4dc-e737-4d48-a041-a715726e9833)


Such data would be harmful to any model being trained on the earlier mentioned tasks (speech decoding)

# LDS
LDS was used to differentiate between good and bad electrodes. The Linear Dynamical System allows us to calculate the log likelihood over the observations, and this is the criteria used to differentiate between good and bad electrodes. The Linear Dynamical System was trained on data only from good electrodes. Now if inference is done on the LDS with good electrode data, we should get a high log likelihood as compared to doing inference on the LDS with bad electrode data.

Note: All the code is my own except for the notch_filter function in data.py