Neuromorphic computing facilitates deep brain-machine fusion for high-performance neuroprosthesis

Yu Qi, Jiajun Chen, Yueming WangFrontiers in Neuroscience 2023 · DOI

Abstract

Brain-machine interfaces (BMI) have developed rapidly in recent years, but still face critical issues such as accuracy and stability. Ideally, a BMI system would be an implantable neuroprosthesis that would be tightly connected and integrated into the brain. However, the heterogeneity of brains and machines hinders deep fusion between the two. Neuromorphic computing models, which mimic the structure and mechanism of biological nervous systems, present a promising approach to developing high-performance neuroprosthesis. The biologically plausible property of neuromorphic models enables homogeneous information representation and computation in the form of discrete spikes between the brain and the machine, promoting deep brain-machine fusion and bringing new breakthroughs for high-performance and long-term usable BMI systems. Furthermore, neuromorphic models can be computed at ultra-low energy costs and thus are suitable for brain-implantable neuroprosthesis devices. The intersection of neuromorphic computing and BMI has great potential to lead the development of reliable, low-power implantable BMI devices and advance the development and application of BMI.

The Challenge: Bridging the Representation Gap

Despite rapid progress in brain-machine interfaces (BMIs)—such as those used to control robotic arms or synthesize speech—two major hurdles remain: heterogeneity and power consumption.

Biological brains communicate via discrete electrical pulses called "spikes." Conversely, traditional computing machines process information using continuous real-valued vectors. To bridge this, current BMIs typically "bin" spikes into time intervals, losing the precise timing and order of neural activity. This mismatch limits the accuracy and stability of the interface. Furthermore, brain signals are non-stationary; they change over time. Most current BMIs require daily recalibration because they lack the "co-adaptation" necessary to evolve alongside the brain.

The Neuromorphic Solution

Neuromorphic computing offers a shift toward a foundational architecture that aligns the machine's computational language with the biological nervous system. By utilizing Spiking Neural Networks (SNNs), researchers can match the brain's native "data format."

Deep brain-machine fusion through spike-based interaction and Hebbian co-learning.

Key components of this approach include:

  • Homogeneous Information Transfer: Moving away from "binned" values to direct spike-based interaction preserves precise temporal information.
  • Biologically Plausible Learning: By implementing Hebbian rules and Spike-Timing-Dependent Plasticity (STDP), the machine can learn in a way that mirrors biological synaptic changes, solving the issue of "catastrophic forgetting" common in traditional AI.
  • Event-Driven Hardware: Utilizing neuromorphic chips (such as IBM’s TrueNorth or Intel’s Loihi) allows for asynchronous processing. These chips only consume significant power when a spike occurs, rather than running a constant clock.

Overcoming the Power Wall

High-performance neural decoding is traditionally computationally expensive. Existing wireless implantable devices often struggle to last more than 1–2 days on a single charge. Because neuromorphic systems are event-driven, they are incredibly efficient. A neuromorphic chip can operate at $10^{-3}$ to $10^{-1}$ Watts, whereas a comparable server-class CPU consumes around $10^2$ Watts. This efficiency is critical for creating fully implantable, long-term clinical devices.

Conceptual illustration of an implantable neuroprosthesis integrated into the biological brain.

Key Results and Proof of Concept

Synthesized findings from several landmark studies highlight the potential of this fusion:

  • High Accuracy: Spiking neural network decoders for real-time cursor control have achieved success rates of over 94% that remained stable for extended periods.
  • Superior Learning: Neuromorphic olfactory circuits have demonstrated "one-shot learning," outperforming traditional machine learning in recognizing odors from limited neural data.
  • Efficiency Gains: Neuromorphic hardware has shown the potential to extend the battery life of implantable devices from days to potentially much longer durations, facilitating the move toward "fully implantable" wireless neuroprosthetics.

Limitations and Future Directions

While the potential is vast, the intersection of neuromorphic computing and BMI is still in its infancy. Most evidence is currently based on offline simulations or theoretical models. Significant engineering challenges remain in developing fully implantable, medical-grade neuromorphic hardware that can be safely integrated into the human cranium. Future research must focus on "deep fusion" where the brain and the machine adapt to each other simultaneously, treating the neuroprosthesis as a seamless extension of the biological brain.

Summary updated

Abstract

Brain-machine interfaces (BMI) have developed rapidly in recent years, but still face critical issues such as accuracy and stability. Ideally, a BMI system would be an implantable neuroprosthesis that would be tightly connected and integrated into the brain. However, the heterogeneity of brains and machines hinders deep fusion between the two. Neuromorphic computing models, which mimic the structure and mechanism of biological nervous systems, present a promising approach to developing high-performance neuroprosthesis. The biologically plausible property of neuromorphic models enables homogeneous information representation and computation in the form of discrete spikes between the brain and the machine, promoting deep brain-machine fusion and bringing new breakthroughs for high-performance and long-term usable BMI systems. Furthermore, neuromorphic models can be computed at ultra-low energy costs and thus are suitable for brain-implantable neuroprosthesis devices. The intersection of neuromorphic computing and BMI has great potential to lead the development of reliable, low-power implantable BMI devices and advance the development and application of BMI.
BibTeX
@inproceedings{qi2023neuromorphic,
  author = {Yu Qi and Jiajun Chen and Yueming Wang},
  title = {{Neuromorphic computing facilitates deep brain-machine fusion for high-performance neuroprosthesis}},
  booktitle = {Frontiers in Neuroscience},
  year = {2023},
  doi = {10.3389/fnins.2023.1153985},
}
Figure