A neural decision forest scheme with application to EMG gesture classification

Herein presents a novel neural decision forest by bootstrap aggregating randomized neural decision trees with the purpose of untangling signal data representations and performing heirarchical classification. The individual trees in a neural decision forest embed randomized neural networks (rNet) with 1-hidden layer acting as bottlenecks and hard gating functions. The rNets are trained using a multi-class exchange method with l
₂
-regularization to sparsify the networks. By applying a projection map prior to each rNet input layer, a novel method to randomly window a vectorized tensor of electromyogram (EMG) signals recorded from hand gestures is attained. These random windows show the promise of greatly simplifying and generalizing the process of feature extraction. Neural decision forests prevent overfitting and the inability of single trees to generalize to new variations by bootstrap sampling the training set, controling the rNet convergence rates, and increasing the forest size. The key findings in three gesture classification experiments are: i) neural decision forests with random windows outperform the random forest algorithm in terms of accuracy measures in personalized gestures by as much as 20% with equal tree sizes, ii) neural decision forests outperform random forests in intra-subject classification discriminating 27 gestures on 4 out of 5 people, and iii) neural decision forests are competitive with random forests in discriminating 7 gestures recorded with a low sampling rate.