This repo includes multiple distinct project.
4-class motor imagery
What I did and/or learned
- Added comments as I read and tried to understand the (very) complicated signal processing
- Tested my ensemble of ML models with and without PCA
- I keep only 2% of the original data and still got extremely competitive results.
- without PCA (all data), I outperformed the authors’ Accuracy: 87.9% vs {86.8%, 85%}, Used a
EnsembleVoteClassifier
Mindbigdata, the "MNIST" of Brain Digits; Given the brain signal(s) of 2 seconds each, captured with the stimulus of seeing a digit (from 0 to 9) and thinking about it, determine what the digit is. Offical site
I really wanted to get this working, I wanted to achieve at least 25% accuracy. Unfortunately I have not scene any team or paper achieve more then about 12.5% accuracy. I'm not sure if its even possible to do better.
-
I did learn a lot about Digital signal processing and EEG data. I one attempt, as I recall, I only keep 52hz & 55hz data... a lot better the idiotic approach of blindly throwing LSTMs at the problem.
-
I also learned how to make Bidirectional LSTMs & ConVLSTMs, and combined the two
Bidirectional(ConVLSTMs(..))
-
My learning outcome was an increased understanding of sequential models, which I applied to CMPT 419, in which I co-authored "Predicting the S&P 500 with LTMs and GloVe". Poster
This dataset is a pre-processed and re-structured/reshaped version of a very commonly used dataset featuring epileptic seizure detection.
The point of this project was
- Give a sanity test where I was faces with marginal results after 15 hours of attempting Mindbigdata, the "MNIST" of Brain Digits
- There was massive code reuse, I just wanted to see if my 3D dataset building and LSTM model was working as intended
3rd-year Cognitive science student at Simon Fraser University. Interested in Machine learning, Computational Data science, Finance, Anomaly detection, and Embedded systems for Automation.
Research Interests: Applications of Qualitative Analytics in Human behavior
...and Agent behavior, however, that’s just sci-fi... for now. This is useful for side-channel attacks on reinforcement learners
(past) VP of Cognitive Science Student Society. Member of the Robot Soccer Club, and Finance club.
Pull requests are welcome.
MNIST brain Data is under the Database Contents License (DbCL) v1.0
Stuff that is mine to License, and License-able has the Unlicense. Do with my work as you wish. Internship offers are the recommended attribution :)