We have collected 50 GB worth of raw data. Finally, now comes the most exciting part of the project that we all have been waiting for – building the machine learning model.
Before we start generating images with our brain waves, we must first be able to read and classify them. To build our classifier, we used a Recurrent Neural Network comprising of 3 layers. A Long-Short-Term-Memory layer (LSTM) followed by 0.6 Dropout, followed by a Dense Layer, finally followed by a Classification (Softmax) Layer. We only arrived at this architecture after much configuration and when it showed the highest test accuracy.
Thankfully, the company we were attached to, China Comservice, had good GPUs and servers for us to train our model. It had reduced our training time from 1 hour to just 10 mins.
To understand the complexity of this classification task, take a look at the 2 images below. After looking at the strips that represent Airplanes and those that represent Cars, could you tell me which class the next strip of EEG falls under?
Bear in mind that there are only 2 classes here, and even with random guessing, you may be able to classify these images correctly half of the time.
However, with 54 classes, the probability of correctly guessing the class is only 1.85%.
I’m proud to say that our Classification Model can correctly classify the data to a test accuracy of 80.6%.