Architectural Exploration of Hybrid Neural Decoders for Neuromorphic Implantable BMI (2505.05983v1)

Abstract: This work presents an efficient decoding pipeline for neuromorphic implantable brain-machine interfaces (Neu-iBMI), leveraging sparse neural event data from an event-based neural sensing scheme. We introduce a tunable event filter (EvFilter), which also functions as a spike detector (EvFilter-SPD), significantly reducing the number of events processed for decoding by 192X and 554X, respectively. The proposed pipeline achieves high decoding performance, up to R^2=0.73, with ANN- and SNN-based decoders, eliminating the need for signal recovery, spike detection, or sorting, commonly performed in conventional iBMI systems. The SNN-Decoder reduces computations and memory required by 5-23X compared to NN-, and LSTM-Decoders, while the ST-NN-Decoder delivers similar performance to an LSTM-Decoder requiring 2.5X fewer resources. This streamlined approach significantly reduces computational and memory demands, making it ideal for low-power, on-implant, or wearable iBMIs.

Overview of Hybrid Neural Decoders for Neuromorphic Implantable BMIs

The paper under review presents an advanced decoding pipeline tailored for neuromorphic implantable brain-machine interfaces (Neu-iBMIs), which effectively leverages sparse neural event data derived from event-based neural sensing techniques. The central innovation introduced is a tunable event filter (EvFilter), which demonstrates dual functionality, both as a generic event filter and as a spike detector (EvFilter-SPD). This component significantly reduces the number of events needed for processing, yielding compression ratios of $192\times$ and $554\times$ for the respective functions. The pipeline's decoding performance reaches $\mathrm{R^2}=0.73$ using both artificial neural networks (ANN) and spiking neural networks (SNN) decoders. These numerical results underscore the efficiency and efficacy of the proposed approach.

Key Contributions

The authors contribute several innovative techniques to the field of neuromorphic computing and brain-machine interfacing:

1. Event Filtering: The EvFilter can be adjusted to suppress background events or operate as a spike detector. The paper details that the EvFilter reduces the number of neural events in the data stream by a significant margin, which is crucial for reducing computational load and energy consumption in wearable or implantable devices.
2. Hybrid Neural Decoders: The paper explores varied decoding architectures such as shallow multi-layer perceptron (MLP) networks, event-based SNN decoders, and comparison against long short-term memory (LSTM) decoders. The SNN-Decoder notably reduces computational and memory requirements by $5-23\times$ compared to traditional neural networks and LSTM-based methods while maintaining high predictive performance.
3. Architectural Innovation for Low-Power Systems: The reduced computational demand of the SNN-based approach, when integrated with EvFilter, showcases the potential for implementing these decoders on low-power, on-implant, and wearable iBMI systems.

Numerical Validation

Robust validation is provided through the use of a primate dataset of hand-reaching tasks, demonstrating the application of the hybrid decoders. The high $\mathrm{R^2}$ scores, with a peak of 0.73 for LSTM decoders, attest to the model's ability to effectively capture the relationship between neural event data and intended motor actions. The superior performance of the proposed models over traditional methods further validates the approach, showing marked improvements over linear regression and other established decoding algorithms like the Kalman filter.

Implications for Future AI Developments

This research holds significant implications for both practical applications and theoretical advancements in neurotechnology:

• Practical Implications: The reduced computational and memory requirements make these decoders feasible for deployment in real-time, without the need for bulky computing hardware. This is particularly advantageous for neuromorphic processors and edge computing devices, which operate under strict power and space constraints.
• Theoretical Implications: The integration of event-based data processing with spiking neural networks suggests new avenues for improving the efficiency of neural decoding systems. The hybrid approach can potentially redefine how event-based data is harnessed for high-performance computing tasks.
• Future Prospects: As neurotechnology progresses towards the development of more sophisticated, high-channel count iBMI systems, the techniques outlined in this paper could form the basis for next-generation interfaces that are both efficient and scalable.

In summary, this paper outlines significant advancements in decoder technology for Neu-iBMIs, effectively balancing performance with computational efficiency. The innovations presented have the potential to enhance future neuroprosthetics and neuromodulation technologies by providing more refined control mechanisms, thereby broadening the scope and applicability of implantable BMI systems.