| name | dynamic-neural-manifolds-control |
| description | Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware. Uses sensory inputs to modulate heterogeneous inhibition, gain, and transient currents, driving rapid subspace rotations to switch between behaviors and fine-grained trajectory control within them. |
| version | 1.0.0 |
| author | Hermes Agent |
| license | MIT |
| metadata | {"hermes":{"tags":["neuroscience","neuromorphic","dynamic-manifolds","closed-loop-control","spiking-neural-networks","spinnaker2","robotic-navigation","explainable-ai"],"category":"ai_collection","arxiv_id":"2607.07373","arxiv_url":"https://arxiv.org/abs/2607.07373","published":"2026-07-08","authors":["Oskar von Seeler","Christian Tetzlaff","Andrew Lehr"],"categories":["cs.NE"],"trigger_words":["dynamic neural manifolds","closed-loop control","neuromorphic hardware","spinnaker","subspace rotation","manifold geometry","behavior switching","trajectory control","spiking network","heterogeneous inhibition"]}} |
| created | 2026-07-12 |
| updated | 2026-07-12 |
Dynamic Neural Manifolds for Flexible Closed-Loop Control on Neuromorphic Hardware
arXiv: 2607.07373 | Published: 2026-07-08 | Authors: Oskar von Seeler, Christian Tetzlaff, Andrew Lehr
Core Thesis
Sequential neural activity in biological circuits evolves along dynamic, low-dimensional manifolds to enable flexible behavior. This paper extends the dynamic neural manifold framework to neuromorphic engineering, implementing it on the SpiNNaker 2 chip for real-time, closed-loop control.
By allowing sensory inputs to modulate heterogeneous inhibition, gain, and transient currents, the architecture drives:
- Rapid subspace rotations to switch between behaviors
- Fine-grained trajectory control within subspaces
Key Concepts
Dynamic Neural Manifolds
- Biological neural activity doesn't explore the full high-dimensional state space; it evolves along low-dimensional manifolds
- These manifolds are dynamic — they can be reshaped by external inputs
- Specific circuit mechanisms link manifold geometry to computational function
Sensory Modulation Mechanisms
Three types of sensory input modulation enable manifold control:
- Heterogeneous inhibition: Different neurons receive different levels of inhibitory input, reshaping the manifold's curvature
- Gain modulation: Multiplicative scaling of neuronal responses changes the manifold's scale and direction
- Transient currents: Brief current injections cause rapid manifold rotations
Subspace Rotations
When the agent needs to switch behaviors (e.g., from "explore" to "avoid"), sensory inputs trigger a rotation of the active subspace, redirecting the trajectory to a different behavioral attractor.
Implementation on SpiNNaker 2
Architecture
Sensory Input → [Inhibition Modulator, Gain Modulator, Transient Current Injector]
↓
Spiking Network (with dynamic manifold structure)
↓
Motor Output → Robotic Agent
↓
Sensory Feedback (closed loop)
Key Design Decisions
- Real-time operation: SpiNNaker 2 enables real-time spike processing
- Parameterizable manifolds: Circuit mechanisms are explicitly tied to manifold geometry
