Reinforcement learning (RL)-based decoders in brain-machine interfaces (BMIs) interpret dynamic neural activity without patients' real limb movements. In conventional RL, the goal state is selected by the user or defined by the physics of the problem, and the decoder finds an optimal policy essentially by assigning credit over time, which is normally very time-consuming. However, BMI tasks require finding a good policy in very few trials, which impose a limit on the complexity of the tasks that can be learned before the animal quits. Therefore, this paper explores the possibility of letting the agent infer potential goals through actions over space with multiple objects, using the instantaneous reward to assign credit spatially. A previous method, attention-gated RL employs a multilayer perceptron trained with backpropagation, but it is prone to local minima entrapment. We propose a quantized attention-gated kernel RL (QAGKRL) to avoid the local minima adaptation in spatial credit assignment and sparsify the network topology. The experimental results show that the QAGKRL achieves higher successful rates and more stable performance, indicating its powerful decoding ability for more sophisticated BMI tasks as required in clinical applications.

Classic brain-machine interface (BMI) approaches decode neural signals from the brain responsible for achieving specific motor movements, which subsequently command prosthetic devices. Brain activities adaptively change during the control of the neuroprosthesis in BMIs, where the alteration of the preferred direction and the modulation of the gain depth are observed. The static neural tuning models have been limited by fixed codes, resulting in a decay of decoding performance over the course of the movement and subsequent instability in motor performance. To achieve stable performance, we propose a dual sequential Monte Carlo adaptive point process method, which models and decodes the gradually changing modulation depth of individual neuron over the course of a movement. We use multichannel neural spike trains from the primary motor cortex of a monkey trained to perform a target pursuit task using a joystick. Our results show that our computational approach successfully tracks the neural modulation depth over time with better goodness-of-fit than classic static neural tuning models, resulting in smaller errors between the true kinematics and the estimations in both simulated and real data. Our novel decoding approach suggests that the brain may employ such strategies to achieve stable motor output, i.e., plastic neural tuning is a feature of neural systems. BMI users may benefit from this adaptive algorithm to achieve more complex and controlled movement outcomes.

Sequential Monte Carlo estimation on point processes has been successfully applied to predict the movement from neural activity. However, there exist some issues along with this method such as the simplified tuning model and the high computational complexity, which may degenerate the decoding performance of motor brain machine interfaces. In this paper, we adopt a general tuning model which takes recent ensemble activity into account. The goodness-of-fit analysis demonstrates that the proposed model can predict the neuronal response more accurately than the one only depending on kinematics. A new sequential Monte Carlo algorithm based on the proposed model is constructed. The algorithm can significantly reduce the root mean square error of decoding results, which decreases 23.6% in position estimation. In addition, we accelerate the decoding speed by implementing the proposed algorithm in a massive parallel manner on GPU. The results demonstrate that the spike trains can be decoded as point process in real time even with 8000 particles or 300 neurons, which is over 10 times faster than the serial implementation. The main contribution of our work is to enable the sequential Monte Carlo algorithm with point process observation to output the movement estimation much faster and more accurately.

Previous decoding algorithms used in brain machine interfaces (BMIs) usually seek a static functional mapping between the spatio-temporal neural activity and behavior and assume that the neural spike statistics do not change over time. However, recent work indicates the significant variance in neural activities, which suggests the nonfeasibility of the stationary assumptions on the neural signal sequences. To track the time-changing neural activity during the nonlinear decoding process, we developed a time-varying approach based on general regression neural network (GRNN) with a dynamic pattern layer. Applied on both simulated neural activity and in vivo BMI data extracted from rat's motor cortex, the proposed method reconstructs the movement signals better than the original GRNN algorithm with static pattern layer, which raises the promise of successfully tracking the time-varying neural activity for BMIs decoding. V

Recently, local field potentials (LFPs) have been successfully used to extract information of arm and hand movement in some brain-machine interfaces (BMIs) studies, which suggested that LFPs would improve the performance of BMI applications because of its long-term stability. However, the performance of LFPs in different frequency bands has not been investigated in decoding hand grasp types. Here, the LFPs from the monkey's dorsal premotor cortices were collected by microelectrode array when monkey was performing grip-specific grasp task. A K-nearest neighbor classifier performed on the power spectrum of LFPs was used to decode grasping movements. The decoding powers of LFPs in different frequency bands, channels and trials used for training were also studied. The results show that the broad high frequency band (200-400Hz) LFPs achieved the best performance with classification accuracy reaching over 0.9. It infers that high frequency band LFPs in PMd cortex could be a promising source of control signals in developing functional BMIs for hand grasping.