dynamic-neural-manifolds-neuromorphic-control

name	dynamic-neural-manifolds-neuromorphic-control
category	neuroscience
created	"2026-07-13T00:00:00.000Z"
arxiv_id	2607.07373
description	Dynamic neural manifold architecture for flexible closed-loop control on neuromorphic hardware — mapping spiking activity to low-dimensional manifold trajectories with sensory-modulated geometry for explainable neural computation.
trigger_words	["dynamic neural manifolds","neuromorphic closed-loop control","SpiNNaker neural manifold","manifold geometry neural computation","subspace rotation behavior switching","explainable neuromorphic architecture","ring attractor manifold","bump attractor propagation"]

Dynamic Neural Manifolds for Neuromorphic Closed-Loop Control

Paper Reference

• Title: Dynamic neural manifolds for flexible closed-loop control on neuromorphic hardware
• Authors: Oskar von Seeler, Christian Tetzlaff, Andrew Lehr
• Published: 2026-07-08
• arXiv: 2607.07373
• Categories: cs.NE

Core Concept

Biological neural circuits organize sequential activity along dynamic, low-dimensional manifolds to enable flexible behavior. This paper extends the dynamic manifold framework from biological modeling to neuromorphic engineering, implementing it on the SpiNNaker 2 chip for real-time closed-loop control.

Key Innovations

1. Manifold-Based Neuromorphic Architecture: Spiking networks implement dynamic neural manifolds where behavior is encoded as trajectories through low-dimensional subspaces, making neural computation explainable and parameterizable.

2. Sensory-Modulated Manifold Geometry: Three control channels enable flexible behavior:

   • Heterogeneous inhibition modulation — drives rapid subspace rotations
   • Gain modulation — controls trajectory speed and stability
   • Transient current injection — enables fine-grained trajectory control

3. Real-Time Closed-Loop on SpiNNaker2: First demonstration of dynamic manifold control on actual neuromorphic hardware with sensory feedback loops operating in real-time.

4. Behavioral Switching via Subspace Rotation: Sensory inputs can rapidly reconfigure manifold geometry, enabling the agent to switch between different behavioral modes.

Methodology

Manifold Parameterization

• Sequential neural activity is mapped to trajectories on a low-dimensional manifold
• Circuit mechanisms (inhibition patterns, gain, transient currents) are explicitly linked to manifold geometry features
• This makes the system explainable: you can read behavior from manifold shape

Control Architecture

Sensory Input → Manifold Modulation → Spiking Activity → Motor Output
                    ↑                                        ↓
                    └────────── Feedback Loop ──────────────┘

1. Sensory encoding maps environmental state to manifold modulation parameters
2. Manifold modulation rotates subspaces and adjusts trajectory parameters
3. Spiking dynamics evolve along the reconfigured manifold
4. Readout extracts motor commands from population activity
5. Closed-loop feedback continuously updates manifold geometry

Validation

• Robotic agent navigation through a maze
• Agent uses sensory feedback to dynamically reconfigure its manifold geometry
• Demonstrates both behavioral switching (subspace rotation) and fine-grained control (trajectory modulation)

Implementation Patterns

For Neuromorphic Systems

• Use heterogeneous inhibitory connectivity to enable subspace rotation
• Implement gain modulation as multiplicative scaling of neuronal inputs
• Add transient current injection for rapid state transitions
• Design readout layers that project population activity to behavior space

For Biological Modeling

• Map circuit-level mechanisms (inhibition, gain, transients) to manifold geometry
• Use manifold analysis to understand how biological circuits achieve flexibility
• Validate by comparing simulated manifold trajectories to neural recordings

Pitfalls

1. Manifold Dimensionality: Too low → insufficient expressivity; too high → loss of explainability. Must balance based on behavioral complexity.
2. Temporal Scales: Subspace rotations must be faster than within-manifold trajectory dynamics, otherwise the controller can't switch behaviors fast enough.
3. Hardware Constraints: SpiNNaker2 has limited precision for current injection; discretization effects can distort manifold geometry.
4. Sensory Encoding Quality: Poor sensory→manifold mapping leads to unstable closed-loop behavior.

Applications

• Neuromorphic robotics: Real-time adaptive control on edge hardware
• Brain-computer interfaces: Explainable decoding of motor intentions
• Neuroscience experiments: Testable predictions about how manifold geometry relates to behavior
• Explainable AI: Understanding how neural circuits implement computation through geometric lens

Related Skills

• spiking-neural-network-analysis
• neuromorphic-artificial-consciousness
• brain-network-controllability