Unsupervised, piecewise linear decoding enables an accurate prediction of muscle activity in a multi-task brain...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Intracortical brain-computer interfaces (iBCIs) require decoders for translating neural activity into commands.
• Nonlinearity in neural signals poses a significant challenge for developing effective global iBCI decoders.
• Seamless task and context transitions are crucial for enhancing iBCI user experience.

Purpose of the Study:

• To develop an unsupervised method for creating piecewise linear decoders for iBCIs.
• To address the challenges posed by neural signal nonlinearity in iBCI decoding.
• To improve iBCI performance across diverse tasks and contexts.

Main Methods:

• Utilized an unsupervised approach based on low-dimensional neural manifold structure.
• Recorded neural signals from motor cortex and upper limb muscle electromyographs (EMGs) in monkeys performing various tasks.
• Applied dimensionality reduction and unsupervised clustering to identify neural manifold structures, fitting a linear decoder for each cluster.

Main Results:

• Identified distinct clusters in neural manifolds corresponding to different tasks or task sub-phases.
• Piecewise decoding performance improved with an increasing number of clusters, eventually plateauing.
• A piecewise linear decoder with just two clusters outperformed a global linear decoder and even a global recurrent neural network decoder.

Conclusions:

• Introduced a computationally efficient solution for iBCI decoders applicable to a wide range of tasks.
• Demonstrated that piecewise linear decoders can effectively approximate the nonlinearity between neural activity and motor outputs.
• The findings enhance understanding of neural manifold structures in the motor cortex for improved iBCI development.