Figure 1: Classification accuracies where we transfer all other moths’ data to a specific target moth’s feature space. The target moths are listed on the X-axis. Our method gives as much as 300% improvement over HiWA.
To this end, we have developed a Restricted Boltzmann Machine trained based on Fisher Divergence (RBM-FD) for transferring one subject’s feature space to another (referred to as target). We have derived a closed form of the gradient of the loss of the RBM-FD that can be explicitly computed in real-time. This enables efficient training using the stochastic gradient descent algorithm. We test the transfer learning capability of our method on the classification of neuro-muscular recordings of spike trains from the ten muscles. The output of the classifier is the visual stimulus corresponding to the neuro-muscular recordings. Subject-to-subject variation clearly exists among the recordings and a subject specific classifier performs poorly on other subjects’ data. Using the FD-RBM, we transfer the data from a source subject to the target subject feature space and use the target subject classifier for decoding. Our results improve on the performance of the State-of-the-art methods (e.g. HiWA) by 300%.