Feedback for reinforcement learning based brain-machine interfaces using confidence metrics

Neural Decoders Using Reinforcement Learning in Brain Machine Interfaces: A Technical Review

Creating flexible and robust brain machine interfaces (BMIs) is currently a popular topic of research that has been explored for decades in medicine, engineering, commercial, and machine-learning communities. In particular, the use of techniques using reinforcement learning (RL) has demonstrated impressive results but is under-represented in the BMI community. To shine more light on this promising relationship, this article aims to provide an exhaustive review of RL's applications to BMIs. Our primary focus in this review is to provide a technical summary of various algorithms used in RL-based BMIs to decode neural intention, without emphasizing preprocessing techniques on the neural signals and reward modeling for RL. We first organize the literature based on the type of RL methods used for neural decoding, and then each algorithm's learning strategy is explained along with its application in BMIs. A comparative analysis highlighting the similarities and uniqueness among neural decoders is provided. Finally, we end this review with a discussion about the current stage of RLBMIs including their limitations and promising directions for future research.

Kernel Temporal Differences for EEG-based Reinforcement Learning Brain Machine Interfaces

Kernel temporal differences (KTD) (λ) algorithm integrated in Q-learning (Q-KTD) has shown its applicability and feasibility for reinforcement learning brain machine interfaces (RLBMIs). RLBMI with its unique learning strategy based on trial-error allows continuous learning and adaptation in BMIs. Q-KTD has shown good performance in both open and closed-loop experiments for finding a proper mapping from neural intention to control commands of an external device. However, previous studies have been limited to intracortical BMIs where monkey's firing rates from primary motor cortex were used as inputs to the neural decoder. This study provides the first attempt to investigate Q-KTD algorithm's applicability in EEG-based RLBMIs. Two different publicly available EEG data sets are considered, we refer to them as Data set A and Data set B. EEG motor imagery tasks are integrated in a single step center-out reaching task, and we observe the open-loop RLBMI experiments reach 100% average success rates after sufficient learning experience. Data set A converges after approximately 20 epochs for raw features and Data set B shows convergence after approximately 40 epochs for both raw and Fourier transform features. Although there still exist challenges to overcome in EEG-based RLBMI using Q-KTD, including increasing the learning speed, and optimization of a continuously growing number of kernel units, the results encourage further investigation of Q-KTD in closed-loop RLBMIs using EEG. Clinical Relevance- This study supports feasibility of noninvasive EEG-based RLBMI implementations and addresses benefits and challenges of RLBMI using EEG.

The value-complexity trade-off for reinforcement learning based brain-computer interfaces

OBJECTIVE

One of the recent developments in the field of brain-computer interfaces (BCI) is the reinforcement learning (RL) based BCI paradigm, which uses neural error responses as the reward feedback on the agent's action. While having several advantages over motor imagery based BCI, the reliability of RL-BCI is critically dependent on the decoding accuracy of noisy neural error signals. A principled method is needed to optimally handle this inherent noise under general conditions.

APPROACH

By determining a trade-off between the expected value and the informational cost of policies, the info-RL (IRL) algorithm provides optimal low-complexity policies, which are robust under noisy reward conditions and achieve the maximal obtainable value. In this work we utilize the IRL algorithm to characterize the maximal obtainable value under different noise levels, which in turn is used to extract the optimal robust policy for each noise level.

MAIN RESULTS

Our simulation results of a setting with Gaussian noise show that the complexity level of the optimal policy is dependent on the reward magnitude but not on the reward variance, whereas the variance determines whether a lower complexity solution is favorable or not. We show how this analysis can be utilized to select optimal robust policies for an RL-BCI and demonstrate its use on EEG data.

SIGNIFICANCE

We propose here a principled method to determine the optimal policy complexity of an RL problem with a noisy reward, which we argue is particularly useful for RL-based BCI paradigms. This framework may be used to minimize initial training time and allow for a more dynamic and robust shared control between the agent and the operator under different conditions.

Noise-Correlation Is Modulated by Reward Expectation in the Primary Motor Cortex Bilaterally During Manual and Observational Tasks in Primates

Reward modulation is represented in the motor cortex (M1) and could be used to implement more accurate decoding models to improve brain-computer interfaces (BCIs; Zhao et al., 2018). Analyzing trial-to-trial noise-correlations between neural units in the presence of rewarding (R) and non-rewarding (NR) stimuli adds to our understanding of cortical network dynamics. We utilized Pearson's correlation coefficient to measure shared variability between simultaneously recorded units (32-112) and found significantly higher noise-correlation and positive correlation between the populations' signal- and noise-correlation during NR trials as compared to R trials. This pattern is evident in data from two non-human primates (NHPs) during single-target center out reaching tasks, both manual and action observation versions. We conducted a mean matched noise-correlation analysis to decouple known interactions between event-triggered firing rate changes and neural correlations. Isolated reward discriminatory units demonstrated stronger correlational changes than units unresponsive to reward firing rate modulation, however, the qualitative response was similar, indicating correlational changes within the network as a whole can serve as another information channel to be exploited by BCIs that track the underlying cortical state, such as reward expectation, or attentional modulation. Reward expectation and attention in return can be utilized with reinforcement learning (RL) towards autonomous BCI updating.

Task Learning Over Multi-Day Recording via Internally Rewarded Reinforcement Learning Based Brain Machine Interfaces

Autonomous brain machine interfaces (BMIs) aim to enable paralyzed people to self-evaluate their movement intention to control external devices. Previous reinforcement learning (RL)-based decoders interpret the mapping between neural activity and movements using the external reward for well-trained subjects, and have not investigated the task learning procedure. The brain has developed a learning mechanism to identify the correct actions that lead to rewards in the new task. This internal guidance can be utilized to replace the external reference to advance BMIs as an autonomous system. In this study, we propose to build an internally rewarded reinforcement learning-based BMI framework using the multi-site recording to demonstrate the autonomous learning ability of the BMI decoder on the new task. We test the model on the neural data collected over multiple days while the rats were learning a new lever discrimination task. The primary motor cortex (M1) and medial prefrontal cortex (mPFC) spikes are interpreted by the proposed RL framework into the discrete lever press actions. The neural activity of the mPFC post the action duration is interpreted as the internal reward information, where a support vector machine is implemented to classify the reward vs. non-reward trials with a high accuracy of 87.5% across subjects. This internal reward is used to replace the external water reward to update the decoder, which is able to adapt to the nonstationary neural activity during subject learning. The multi-cortical recording allows us to take in more cortical recordings as input and uses internal critics to guide the decoder learning. Comparing with the classic decoder using M1 activity as the only input and external guidance, the proposed system with multi-cortical recordings shows a better decoding accuracy. More importantly, our internally rewarded decoder demonstrates the autonomous learning ability on the new task as the decoder successfully addresses the time-variant neural patterns while subjects are learning, and works asymptotically as the subjects' behavioral learning progresses. It reveals the potential of endowing BMIs with autonomous task learning ability in the RL framework.

Deep Learning Based Inter-subject Continuous Decoding of Motor Imagery for Practical Brain-Computer Interfaces

Inter-subject transfer learning is a long-standing problem in brain-computer interfaces (BCIs) and has not yet been fully realized due to high inter-subject variability in the brain signals related to motor imagery (MI). The recent success of deep learning-based algorithms in classifying different brain signals warrants further exploration to determine whether it is feasible for the inter-subject continuous decoding of MI signals to provide contingent neurofeedback which is important for neurorehabilitative BCI designs. In this paper, we have shown how a convolutional neural network (CNN) based deep learning framework can be used for inter-subject continuous decoding of MI related electroencephalographic (EEG) signals using the novel concept of Mega Blocks for adapting the network against inter-subject variabilities. These Mega Blocks have the capacity to repeat a specific architectural block several times such as one or more convolutional layers in a single Mega Block. The parameters of such Mega Blocks can be optimized using Bayesian hyperparameter optimization. The results, obtained on the publicly available BCI competition IV-2b dataset, yields an average inter-subject continuous decoding accuracy of 71.49% (κ = 0.42) and 70.84% (κ = 0.42) for two different training methods such as adaptive moment estimation (Adam) and stochastic gradient descent (SGDM), respectively, in 7 out of 9 subjects. Our results show for the first time that it is feasible to use CNN based architectures for inter-subject continuous decoding with a sufficient level of accuracy for developing calibration-free MI-BCIs for practical purposes.

A New Frontier: The Convergence of Nanotechnology, Brain Machine Interfaces, and Artificial Intelligence

A confluence of technological capabilities is creating an opportunity for machine learning and artificial intelligence (AI) to enable "smart" nanoengineered brain machine interfaces (BMI). This new generation of technologies will be able to communicate with the brain in ways that support contextual learning and adaptation to changing functional requirements. This applies to both invasive technologies aimed at restoring neurological function, as in the case of neural prosthesis, as well as non-invasive technologies enabled by signals such as electroencephalograph (EEG). Advances in computation, hardware, and algorithms that learn and adapt in a contextually dependent way will be able to leverage the capabilities that nanoengineering offers the design and functionality of BMI. We explore the enabling capabilities that these devices may exhibit, why they matter, and the state of the technologies necessary to build them. We also discuss a number of open technical challenges and problems that will need to be solved in order to achieve this.