Generating Artificial Neural Signals to Improve the Training of Reinforcement Learning Brain-Computer Interfaces

With the research on reinforcement learning brain-computer interfaces (RLBCI) advancing towards more complex algorithms, the amount of data necessary to train these deep networks must increase tenfold. Unfortunately, it is complicated to get consistent data from non-human primates (NHP); some days the animal might not work, which might impact their work schedule. The high price of maintaining an NHP colony also limits animal research for small research laboratories. A solution for these issues is the use of synthetically generated data neurons where a strong representation of the neural state must be used for high fidelity to the brain responses necessary for the RLBCI algorithms. Recent research around autoencoder network architecture has allowed us to understand the underlying dynamics of neural signals and how these signals can be used to better simulate artificial data. The first aim of this dissertation proposes that the dynamical system obtained from the NHP’s primary motor cortex not only contains important information about the reward landscape, but this information can be transferred successfully to synthetic data. Aim two delves into the discovery of simultaneous kinematic and reward data embedded in the latent space of the motor cortex which can also be embedded into synthetic units. With the results of this work, we have proven the existence of significant differences between reward and non-reward conditions as well as differences on kinematic and reward signals present in the motor cortex’s latent system. We also demonstrated that the obtained latent representation outperforms the more traditional methods of dimensionality reduction when used in a classification task. Moreover, the synthetic data we created, behaves similar to the real data and carries the observed differences between conditions, proven by the performance of the artificial data by a neural network classification. The findings presented in this work are strong evidence for becoming stepstones for the further development of synthetic data that mimics as much as possible real data. Therefore, one day we might not need the use of animals in neuroscience research.