Recurrent Neural Networks with Weighting Loss for Early prediction of Hand Movements

We propose in this work an approach for early prediction of hand movements using recurrent neural networks (RNNs) and a novel weighting loss. The proposed loss function leverages the outputs of an RNN at different time steps and weights their contributions to the final loss linearly over time steps. Additionally, a sample weighting scheme also constitutes a part of the weighting loss to deal with the scarcity of the samples where a change of hand movements takes place. The experiments conducted with the Ninapro database reveal that our proposed approach not only improves the performance in the early prediction task but also obtains state of the art results in classification of hand movements. These results are especially promising for the amputees.