dynamic-neural-manifolds-neuromorphic

name	dynamic-neural-manifolds-neuromorphic
description	Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware. Implements low-dimensional manifold geometry on SpiNNaker 2 chip for real-time robotic control. Activation: neural manifolds, neuromorphic control, spiking networks, manifold geometry, SpiNNaker, closed-loop control, robotic navigation
tags	["neural-manifolds","neuromorphic","spiking-networks","closed-loop-control","spinnaker"]

Core Methodology

Dynamic Neural Manifolds Framework

• Sequential neural activity evolves along low-dimensional dynamic manifolds
• Manifold geometry is parameterizable through circuit mechanisms
• Provides explainable framework for neural computation
• Enables flexible behavior switching via subspace rotations

Key Circuit Mechanisms

Sensory inputs modulate three key parameters:

1. Heterogeneous inhibition - controls subspace geometry
2. Gain modulation - scales activity within manifold
3. Transient currents - drives rapid state transitions

Behavior Switching

• Subspace rotations enable rapid switching between behaviors
• Fine-grained trajectory control within behavioral manifolds
• Sensory feedback dynamically reconfigures manifold geometry

Implementation on SpiNNaker 2

Hardware Architecture

• Real-time closed-loop control on neuromorphic chip
• Event-driven spiking neural networks
• Low-latency sensory-motor integration
• Energy-efficient computation

Validation

• Robotic simulation with maze navigation task
• Agent uses sensory feedback to reconfigure manifolds
• Demonstrates explainable neuromorphic computation
• Substrate for investigating biological neural dynamics

Applications

Neuromorphic Robotics

• Real-time adaptive navigation
• Dynamic behavior switching
• Energy-efficient motor control
• Explainable decision-making

Computational Neuroscience

• Modeling biological neural manifolds
• Understanding sequential activity patterns
• Studying flexible behavior generation
• Bridging neural dynamics and computation

Brain-Inspired AI

• Low-dimensional representation learning
• Dynamic state-space models
• Efficient continual learning
• Robust sensorimotor integration

Implementation Guide

Step 1: Define Manifold Structure

# Identify low-dimensional subspace from neural activity
from sklearn.decomposition import PCA

# Extract principal components of neural activity
pca = PCA(n_components=k)  # k = manifold dimension
manifold_basis = pca.fit_transform(neural_activity)

Step 2: Implement Sensory Modulation

# Sensory input modulates circuit parameters
def modulate_circuit(sensory_input, baseline_params):
    inhibition = baseline_params['inhibition'] * sensory_gain(sensory_input)
    gain = baseline_params['gain'] * (1 + sensory_modulation(sensory_input))
    transient = compute_transient_current(sensory_input)
    return inhibition, gain, transient

Step 3: Drive Subspace Rotations

# Rotate manifold to switch behaviors
def rotate_manifold(current_state, target_behavior):
    rotation_matrix = compute_rotation(current_state, target_behavior)
    new_state = rotation_matrix @ current_state
    return new_state

Step 4: Deploy on SpiNNaker 2

• Use PyNN for hardware abstraction
• Configure neural populations with manifold parameters
• Implement sensory-motor loops with low latency
• Monitor neural activity for manifold analysis

Pitfalls

Manifold Dimensionality Selection

Problem: Choosing wrong manifold dimension (k) degrades performance Solution: Use variance explained threshold (e.g., 95%) or cross-validation

Latency Constraints

Problem: Real-time control requires <10ms loop latency Solution: Optimize event routing, use direct hardware access, minimize Python overhead

Stability-Plasticity Balance

Problem: Rapid switching can destabilize learned manifolds Solution: Implement gradual rotation with momentum, use eligibility traces

Verification

Manifold Quality Metrics

• Variance explained by top-k components (>90%)
• Geodesic distance preservation
• Temporal smoothness of trajectories

Control Performance

• Task completion rate
• Switching latency (<100ms for behavior transitions)
• Energy efficiency (spikes per successful action)

Biological Plausibility

• Compare with recorded neural manifold geometry
• Validate subspace rotation dynamics
• Check consistency with experimental data

References

• Paper: arXiv:2607.07373
• Authors: Oskar von Seeler, Christian Tetzlaff, Andrew Lehr
• Hardware: SpiNNaker 2 neuromorphic chip
• Date: July 8, 2026