dynamic-neural-manifolds-snn-control

name	dynamic-neural-manifolds-snn-control
description	Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware — using circuit mechanisms (heterogeneous inhibition, gain modulation, transient currents) as control knobs for manifold geometry, enabling explainable, energy-efficient autonomous behavior. Implemented on SpiNNaker 2 chip with robotic maze navigation validation.
category	neuroscience

Dynamic Neural Manifolds for Flexible Closed-Loop Control

Overview

This skill implements the dynamic neural manifold framework from arXiv:2607.07373v1, which bridges biological neural dynamics with neuromorphic engineering for explainable closed-loop control. The core insight: neural activity in biological circuits evolves along dynamic, low-dimensional manifolds, and specific circuit mechanisms can serve as "control knobs" for manifold geometry.

Key paper: Dynamic neural manifolds for flexible closed-loop control on neuromorphic hardware (von Seeler, Tetzlaff, Lehr — 2026-07-08)

Core Methodology

Neural Manifold Framework

Population activity of N neurons is represented as a trajectory in N-dimensional state space, constrained to low-dimensional manifolds capturing latent task variables. Geometric features map to behavioral execution:

• Subspace rotations: switching between behaviors/movements
• Trajectory speed: adapting movement timing
• Trajectory curvature: controlling execution dynamics

Circuit Control Knobs

Three circuit-level mechanisms provide geometric control:

1. Heterogeneous Inhibition: Facilitates subspace reorientation, allowing sequences to rotate into new hyperplanes to switch behavioral states.

2. Gain Modulation: Regulates sequence propagation speed, providing direct control over neural trajectory velocity.

3. Transient Synaptic Currents: Control trajectory curvature, enabling fine-grained adjustments during behavior execution.

Ring Network Architecture

Control Neurons (speed, shape, selection)
         ↓
    Ring Network (circulant weight matrix, ~50% sparsity)
         ↓
    Downstream Readout Neurons

The ring network generates oscillatory sequential activity on low-dimensional manifolds. Control neurons modulate:

• Propagation speed around the ring
• Width of activity bump
• Activation sparsity

Closed-Loop Implementation

Sensory inputs modulate circuit mechanisms in real-time:

• Environmental cues → heterogeneous inhibition → behavioral state switching
• Sensory feedback → gain modulation → trajectory adjustment
• Task demands → transient currents → execution timing

Implementation Patterns

SpiNNaker 2 Hardware Deployment

# SpiNNaker 2 ring network implementation
# Ring network + readout on-chip, host interface for agent-environment interaction
# Energy-efficient, low-latency substrate for adaptive neuromorphic control

# Key mapping:
# - Circuit mechanisms → SpiNNaker neuron/synapse parameters
# - Sensory inputs → real-time neuromodulation signals
# - Readout → motor commands / action selection

Robotic Navigation Agent

The validated application: two-wheeled robot navigating a maze using:

• Local environmental cues as sensory feedback
• Dynamic manifold reconfiguration for behavior switching (steering vs. jumping)
• Velocity and trajectory adjustment during task execution

Biological Validation

The framework connects to observed biological phenomena:

• Rat spinal cord oscillatory sequences
• Drosophila larvae ganglia dynamics
• Turtle spinal cord patterns
• Mouse medial entorhinal cortex (MEC) sequences

All species show structured dynamic sequences on low-dimensional manifolds, supporting the universality of this computational principle.

Key Advantages

1. Explainability: Internal state is mathematically interpretable — map circuit architecture to geometric features to behavior
2. Energy Efficiency: Neuromorphic hardware enables biological-level efficiency
3. Flexibility: Rapid subspace rotations switch behaviors; fine-grained trajectory control within behaviors
4. Biological Plausibility: Grounded in observed neural dynamics across species

Pitfalls

• Ring network requires circulant weight structure: Non-circulant connectivity breaks the manifold geometry
• Control neuron inputs must be properly scaled: Unbounded inputs can destabilize the manifold
• Hardware mapping needs calibration: SpiNNaker 2 parameters don't map 1:1 to biological parameters
• Readout layer must be designed for the specific task: Generic decoders may not capture manifold-relevant features

Activation Keywords

dynamic neural manifolds, neuromorphic control, SpiNNaker, closed-loop control, neural manifold, ring attractor, geometric neural computation, explainable neuromorphic, behavioral switching, subspace rotation, neural trajectory, oscillatory sequences, brain-inspired control