SUMMARY
This method enables high fidelity decoding of neural activity in an interpretable manner with a small training data set.
The Unmet Need: Measuring neural activity in the brain with a strong signal to noise ratio
- Brain-Machine Interfaces (BMIs) are systems that use signals from the brain to control a computer, robotic arm, or other device. At the core of BMIs are algorithms (“decoders”) that translate signals from the brain into control signals that are used to control an external device.
- Using recorded neural activity for decoders is challenging because current methods can record from only a modest number of the neurons in the brain at one time and because these signals are intrinsically noisy.
- Prior decoding algorithms have leveraged only the instantaneous relationship between neural activity and movement. Recent work has shown that neural populations acts like a “dynamical system”: the current population state strongly predicts the population state a few moments later.
The Proposed Solution: Leverage the dynamical system nature of the brain
The faculty inventors have developed a new algorithm, Decoding using Manifold Neural Dynamics (DeMaND), which leverages knowledge of how the brain acts as a dynamical systems to greatly improve decoding quality and requires less training data. DeMaND uses a two-step procedure to decode signals from neural populations. First, a continuous manifold is generated based on the neural population activity over time and the desired external control signals (e.g., arm/hand movements). This manifold is then gridded to create a discrete approximation to the continuous manifold. Second, to decode new neural data, DeMaND probabilistically infers a position on the manifold grid from binned spike counts of the recorded neurons. At each point in time, it outputs the corresponding values of control signals.
By using interpolation, DeMaND is able to perform few-shot generalization from a small number of training example to decode a large space of control signal values over time. By constraining the population state to a manifold, and assuming dynamics on this manifold, DeMaND is able to more accurate infer population state, and therefore produce more reasonable control signal values over time. DeMaND is well-suited in particular to “motor” applications such as producing control signals for neural prosthetics in patients with tetraplegia, an active area of research and one which several major companies (Neuralink, Meta) are doing development.
ADVANTAGES
ADVANTAGES
- Improved decoding quality
- High fidelity
- Less training data
APPLICATIONS
- Neural Control of Prosthetics
- Brain Interfaces
- Neural Sensors
Proof of concept
Patent Pending
Licensing,Co-development
Matthew Kaufman