Making brain-machine interfaces robust to future neural variability (1610.05872v1)

Abstract: A major hurdle to clinical translation of brain-machine interfaces (BMIs) is that current decoders, which are trained from a small quantity of recent data, become ineffective when neural recording conditions subsequently change. We tested whether a decoder could be made more robust to future neural variability by training it to handle a variety of recording conditions sampled from months of previously collected data as well as synthetic training data perturbations. We developed a new multiplicative recurrent neural network BMI decoder that successfully learned a large variety of neural-to- kinematic mappings and became more robust with larger training datasets. When tested with a non-human primate preclinical BMI model, this decoder was robust under conditions that disabled a state-of-the-art Kalman filter based decoder. These results validate a new BMI strategy in which accumulated data history is effectively harnessed, and may facilitate reliable daily BMI use by reducing decoder retraining downtime.

• The paper proposes a novel multiplicative recurrent neural network (MRNN) decoder trained on extensive historical data to make brain-machine interfaces robust to neural variability, reducing the need for frequent retraining.
• Evaluated in non-human primates, the MRNN decoder significantly outperformed a state-of-the-art Kalman Filter in target acquisition tasks under challenging conditions, including unplanned electrode loss.
• Training with large, augmented datasets across hundreds of days yielded superior offline decoding accuracy (r2 0.81-0.84) and improved closed-loop performance without trading accuracy for robustness.

Robust Brain-Machine Interfaces through Multiplicative Recurrent Neural Networks

This paper addresses a key challenge in the development of clinically viable brain-machine interfaces (BMIs): the degradation of decoder effectiveness due to variability in neural recording conditions over time. The authors propose a solution involving the use of a new multiplicative recurrent neural network (MRNN) decoder, which leverages a large repository of historical neural data to enhance decoder robustness. This decoder is trained with extensive datasets, both actual and artificially perturbed, to anticipate a broad range of neural variabilities, thus reducing the necessity for frequent retraining.

The paper investigates the use of an MRNN, an advanced recurrent neural network architecture, to improve the reliability of BMIs. Unlike traditional linear decoders, which are limited in computational complexity and prone to underfitting when tasked with handling heterogeneous training data, the MRNN can learn complex neural-to-kinematic mappings. This decoder dynamically modifies its recurrent weights based on input, allowing it to build a repertoire of mappings adaptable to varying recording conditions. The MRNN's architecture, originally conceived for character-level LLMing, is adept at managing time-dependent state changes, aligning well with the dynamic nature of neural signals captured during motor tasks.

This research utilized chronic BMI systems implanted in non-human primates to evaluate decoder performance across different recording sessions. The MRNN was assessed for its ability to decode neural signals into hand movement kinematics across hundreds of days, demonstrating superior robustness even against sudden recording condition changes such as the unplanned loss of key electrodes. Notably, under these challenging conditions, the MRNN consistently outperformed a state-of-the-art Feedback Intention Trained Kalman Filter (FIT-KF) in target acquisition tasks.

A prominent aspect of the paper is the incorporation of significant amounts of training data spanning months of recordings. By exploiting this comprehensive data set, the MRNN showed improved offline and closed-loop performance without a trade-off between accuracy and robustness. For instance, the MRNN achieved offline decoding accuracy across recording sessions with r² values of 0.81 and 0.84 for the two subjects, in stark contrast to 0.52-0.57 for the FIT-KF under comparable conditions. The large training data corpus also facilitates the MRNN handling unexpected electrode losses, a typical failure mode in BMI applications.

Training the MRNN involved realistic perturbations to input data, effectively broadening its ability to generalize across unseen recording conditions. This approach, called data augmentation, was particularly beneficial under electrode-dropping scenarios and naturally occurring neural signal variabilities, underscoring the MRNN's resilience.

Implications of this research extend both practically and theoretically. Practically, a robust decoder reduces downtime for BMI users, paving the way for more seamless integration into daily use. Theoretically, it highlights the potential of recurrent neural networks in learning from extensive, varied datasets, opening avenues for robust computational models in dynamic environments.

As neural interface technology advances, future research may explore combining robust MRNN decoders with adaptive decoding frameworks to further enhance system reliability. Additionally, continued exploration of synthetic training perturbations could yield further strides in decoder resilience. This work underscores the viability of a robust BMI strategy using non-linear, computationally powerful decoders and represents a significant step towards practical clinical translation.