The Summit 2013 - Picture by Dan Taylor / Heisenberg Media - www.heisenbergmedia.com

“I think we are about 8 to 10 years away from this being usable by people with no disability”

dashersw

Facebook VP of Engineering Regina Dugan F8 - Picture by Anthony Quintano

dashersw

dashersw

dashersw

dashersw

How it looks now

dashersw

Actually...

Let me show you.

dashersw

How does this black magic work?

100 billion neurons

Chemical reactions of K & Na exchange produce potential difference > "action potential"

Up to 256 electrodes for non-invasive methods

Thousands of electrodes for invasive operations

... and we still don't know how it exactly works.

dashersw

P300 — oddball paradigm

First published by E. Donchin and L.A. Farwell in 1988

2.3 letters / min

Based on randomly flashing letters in a matrix

Rare positive, common negative flashes pique interest

Action potentials fire 300ms after stimulation

dashersw

Separate Black from red

dashersw

BRAIN-MONITOR

Available on GitHub: https://github.com/dashersw/brain-monitor

dashersw

Open-source P300 speller

that still needs some help!

brain-bits

Available on GitHub: https://github.com/dashersw/brain-bits

dashersw

Tech stack

JavaScript all the way up

Native C add-ons for the headset

Node.js for processing

Electron as the wrapper

Vue.js for front-end

brain.js for neural networks

Custom BLDA in JS as another classifier

...and, MATLAB as the server for eig function.

dashersw

Formatting the data for Machine Learning

1. Get raw electrode data
2. Group by each second to get epochs
3. Filter each epoc between 1-12Hz
4. Winsorize data
5. Zero-mean normalization
6. Decimate by 8
7. Apply machine learning
   1. Use brain.js for neural networks
   2. Use a custom algorithm for BLDA