At the beginning of this week, our group did a demo of our project, and we got a lot of useful feedback. Byron suggested including simulated data as a part of our training data. This data would come from taking the ground truth of the signals from each muscle movement and adding noise to it. After receiving that advice, I worked on defining the “ground truth” signals of our data, by analyzing the output of our test data and averaging together the results, while subtracting the noise. Using this ground truth, I superpositioned it with different kinds of additive noise (White/Black/Gaussian/Brownian/Contaminated/Cauchy) to create a diverse set of simulated data for each class. This simulated data can be used for training our model, making it more accurate and fine tined, and more capable of handling signal variations. This will also help us in our early classifications, as it will improve our model, making testing more accurate. Furthermore, with the ground truth distributions, we can use the machine learning trick of MAP to make our ground truth signals a weighted prior distribution. This will add to the complexity of our classifier, which will make it slower, but increase the accuracy.

Next week, I will use all of the data preprocessing that Kayla worked on this week to make and train the classifier. A goal for next week is to have a functional signal classifier, so that we can go home for Thanksgiving break with a working system. Because I spent this week working on something that wasn’t in our schedule, I’ve fallen a bit behind, but I hope to catch up in the next week working with Kayla to make our classifier.