My main project this week included speeding up the EEG prediction model to work in real time and refactoring the production process of the ML modeling process so we can get models to operate quickly and correctly with the right set of inputs feeding into the model for the real-time system. Now it is very quick and efficient to integrate different ML models into the decisioning process in the real-time system. This included many refactoring processes and reorganizing our file system layout. I also added additional features that plot all the inputs against one another so we can observe on the sample data how well a model is able to perform given how linearly separable or inseparable the input features are. For example, attached below is a plot of two features. The feature on the x-axis is the maximum reading of the AF4 electrode for each sample taken and the feature on the y-axis is the ratio of the maximums of AF3 and AF4 subtracted by the mean. These features are selected because they probably correlate with whether a right wink occurred, where we would see a lower maximum ratio, corresponding to a lot of activity in the right node and little in the left node (hence, minimizing the ratio of AF3 and AF4) and this generated plot gives us a rough idea that using the ratio to separate and predict right winking is pretty closely linearly separable. Hence, we have better confidence our model is predicting using good features. I am on track with our team’s progress goals, and plan to implement the double blink and triple blink detection models and a finite state machine to control how and when to apply a wink detection model we have already trained versus applying a double blink detection model. Furthermore, we need some data samples of plain noise to be able to alert our application if it should attempt to classify real-time data, as to not create a bunch of false predictions due to a bad connection.
