membrane-potential-alignment-intracortical-bci

name	membrane-potential-alignment-intracortical-bci
description	Membrane Potential Alignment (MPA) - Test-time adaptation method for spiking neural networks in intracortical brain-computer interfaces. Realigns pretrained decoders to shifted neural recordings by matching membrane potential distributions via KL divergence - computationally efficient for implantable hardware. Activation: test-time adaptation, intracortical BCI, membrane potential alignment, SNN adaptation, neural signal shift, KL divergence matching, unsupervised adaptation.
metadata	{"arxiv_id":"2606.14866","published":"2026-06-12","authors":["Jihun Lee","Sung Woo Park"],"tags":["BCI","spiking-neural-network","test-time-adaptation","membrane-potential","intracortical","KL-divergence","unsupervised","implantable-hardware"]}
license	Complete terms in LICENSE.txt

Membrane Potential Alignment for Intracortical BCI

Overview

Membrane Potential Alignment (MPA) is a test-time adaptation method for spiking neural networks (SNNs) in intracortical brain-computer interfaces. It realigns pretrained decoders to day-to-day neural signal shifts by matching membrane potential distributions via KL divergence - designed to be computationally efficient for implantable hardware.

arXiv: 2606.14866
Authors: Jihun Lee, Sung Woo Park
Published: June 12, 2026
Categories: cs.NE

Core Innovation

Problem: Neural Signal Shifts

Intracortical BCIs suffer from day-to-day neural signal drifts:

• Pretrained decoders degrade over time
• Signal distributions shift across sessions
• Performance drops without retraining

Existing solutions:

• Deep recurrent networks (computationally expensive)
• Adversarial adaptation methods (too heavy for implants)
• Require re-training with labeled data

MPA solution: Lightweight, unsupervised, test-time adaptation.

Key Contributions

1. Membrane Potential Alignment:

   • Aligns internal SNN states to shifted recordings
   • Uses KL divergence for distribution matching
   • No external data or labels required

2. Hardware Efficiency:

   • Computationally lightweight
   • Suitable for implantable devices
   • Avoids expensive recurrent/adversarial operations

3. Test-Time Adaptation:

   • Real-time recalibration during use
   • No offline re-training
   • Continuous performance maintenance

Methodology

Core Concept: MPA Mechanism

Spiking Neural Networks have internal membrane potentials that:

• Encode neural state information
• Reflect signal distribution characteristics
• Change with input signal shifts

MPA approach:

• Compare membrane potential distributions (training vs. current)
• Compute KL divergence mismatch
• Adjust network parameters to realign distributions

Technical Framework

Step 1: Distribution Extraction

# Extract membrane potential statistics
def get_potential_distribution(snn, current_inputs):
    membrane_potentials = []
    for neuron in snn.hidden_layers:
        v_m = neuron.membrane_potential
        membrane_potentials.append(v_m)
    return distribution(membrane_potentials)

Step 2: KL Divergence Computation

# Compute distribution mismatch
def compute_alignment_loss(p_train, p_current):
    kl_divergence = KL(p_current || p_train)
    return kl_divergence

Step 3: Parameter Adjustment

# Realign membrane potentials
def adapt_snn(snn, current_inputs, target_distribution):
    loss = compute_alignment_loss(target_distribution, current_distribution)
    update_parameters(snn, loss)  # Lightweight optimization

Implementation Details

Algorithm:

Input: Pretrained SNN decoder, shifted neural recording batch
Output: Adapted SNN aligned to current signals

1. Forward pass: Get membrane potentials for shifted batch
2. Compute KL divergence with training distribution
3. Gradient descent on alignment loss (few iterations)
4. Update decoder parameters
5. Repeat periodically during operation

Key Properties:

• Unsupervised: No labels needed
• Real-time: Adaptation during inference
• Low-cost: Simple distribution matching

Validation Metrics

1. Decoding Accuracy:

   • Restore performance after shift
   • Compare to pretrained vs. adapted

2. Computational Efficiency:

   • Measure adaptation overhead
   • Ensure implantable hardware feasibility

3. Stability:

   • Consistent performance across days
   • Robust to signal variations

Technical Architecture

Spiking Neural Network Setup

SNN structure:

• Input layer: Neural recording channels
• Hidden layers: Spiking neurons (LIF/IF)
• Output layer: Decoded command signals

Membrane potential dynamics:

dv/dt = -λ*v + I_syn  # Membrane potential evolution
v → spike when v > threshold

Alignment Objective

KL divergence between distributions:

L_align = KL(p_current || p_reference)

Optimization:

• Gradient-based parameter updates
• Few iterations (efficient)
• Preserves pretrained knowledge

Computational Efficiency

Comparison with existing methods:

Method	Adaptation Cost	Suitable for Implants?
Deep recurrent	Heavy (100s of neurons)	No
Adversarial	Very heavy	No
MPA	Lightweight (KL + gradients)	Yes

Applications

Use Cases

1. Intracortical BCI:

   • Real-time decoder adaptation
   • Maintain accuracy across sessions

2. Long-term Implants:

   • Continuous recalibration
   • Patient-independent operation

3. Neural Prosthetics:

   • Robust decoding over months
   • Adapt to signal changes

4. Clinical Deployment:

   • Implantable hardware compatibility
   • Low computational overhead

Integration Patterns

Combine with:

• SNN-based BCI decoders
• Online learning frameworks
• Adaptive control systems

Experimental Results

Key Findings:

1. Performance Recovery:

   • MPA restores ~80-90% of original accuracy after shifts
   • Outperforms naive no-adaptation

2. Hardware Feasibility:

   • Computation time: milliseconds to seconds
   • Memory overhead: minimal

3. Long-term Stability:

   • Maintains performance across multiple sessions
   • Prevents catastrophic degradation

Technical Pitfalls

Common Issues

1. Distribution Estimation:

   • Requires sufficient samples for reliable statistics
   • Small batches may misestimate

2. Over-adaptation:

   • Too many updates can drift from pretrained knowledge
   • Need regularization

3. Signal Variability:

   • Extreme shifts may exceed alignment capacity
   • Require reference distribution updates

4. Hardware Constraints:

   • Gradient computation may exceed budget
   • Need simplified optimization

Solutions

• Use batch statistics with sufficient samples
• Limit adaptation iterations
• Update reference distribution periodically
• Implement lightweight gradient approximations

Activation Keywords

• membrane-potential-alignment, MPA
• test-time-adaptation SNN
• intracortical BCI, intracortical neural decoding
• neural signal shift adaptation
• KL divergence membrane potential
• unsupervised decoder adaptation
• implantable hardware BCI

Related Skills

• test-time-adaptation-benchmark - TTA method comparisons
• snn-learning-survey - SNN training approaches
• cross-subject-eeg-decoding - Cross-subject adaptation
• tta-eeg-foundation-models - TTA for foundation models

References

• arXiv:2606.14866 - MPA original paper
• SNN adaptation surveys
• Intracortical BCI literature
• KL divergence distribution matching

Example Usage

Scenario: Intracortical BCI decoder drift over days

Day 1: Pretrained decoder achieves 95% accuracy
Day 5: Signal shift → accuracy drops to 60%
Apply MPA: Real-time adaptation during operation
Result: Restored to 85-90% accuracy without re-training

Workflow:

1. Deploy pretrained SNN decoder
2. Monitor membrane potential distributions
3. Compute KL divergence when shift detected
4. Run lightweight adaptation (few iterations)
5. Continue decoding with updated parameters