I have a dataset with EMG activations for 8 muscles for several different kinds of movements (
bench-press, etc.) The data has been labelled and segmented into single reps, with each "rep" representing an 8x~100 matrix. (data sampled at 40Hz, 2.5 seconds per rep = approx. 100 samples).
Planning to train a neural network to classify new reps based on this training data. I am unsure about what architectures are common for this type of task (multivariate time series classification). My intuition is that as a rep is performed, a path is traced out through the 8-dimensional space, with the shape and location of the path hopefully being unique to that particular movement. I have a few ideas below:
- High (8) dimensional convolutional neural network with regularization to prevent overfitting. CNN because the shape of the path is likely a strong indicator of class.
- Recurrent neural network, since this is a time-series classification task.
- Some kind of spatial ML model (for ex. SVM), however I am skeptical this could work given issues with time-shifting, etc.
I am implementing everything in tensorflow. What ML-model or NN architecture is best, and is there some existing best-practice for this type of problem? The most unique parts of this problem are:
- Sparse time series
- Multivariate time series
- Relatively small amount of training data (few thousand reps)