dynamic-neural-manifolds-neuromorphic-control

name	dynamic-neural-manifolds-neuromorphic-control
description	Dynamic neural manifolds for flexible closed-loop control on neuromorphic hardware. Implements a ring attractor spiking network on the SpiNNaker 2 chip where sensory-modulated heterogeneous inhibition, multiplicative gain, and transient currents drive rapid subspace rotations and fine-grained trajectory control within low-dimensional neural manifolds. Validated with a robotic maze-navigation simulation. Provides an explainable, neuroscience-grounded framework for mapping world-model plans onto motor control via manifold geometry. Applicable to: neuromorphic control, neural manifolds, ring attractor networks, subspace rotation, closed-loop SNN, SpiNNaker 2, low-dimensional neural dynamics, explainable neuromorphic architectures, behavioral switching.
license	Complete terms in LICENSE.txt
metadata	{"arxiv_id":"2607.07373","published":"2026-07-08","authors":"Oskar von Seeler, Christian Tetzlaff, Andrew Lehr","tags":["neuromorphic-computing","spiking-neural-networks","neural-manifolds","ring-attractor","subspace-rotation","closed-loop-control","spinnaker2","low-dimensional-dynamics","behavioral-switching","explainable-ai"]}

Dynamic Neural Manifolds for Flexible Closed-Loop Control on Neuromorphic Hardware

arXiv: 2607.07373 | Published: 2026-07-08 | Category: cs.NE

Core Contribution

Proposes an explainable parameterization of neural activity as a low-dimensional manifold whose geometry is directly controllable by simple circuit mechanisms, and demonstrates a real-time closed-loop implementation on the SpiNNaker 2 neuromorphic chip. By letting sensory input modulate heterogeneous inhibition, gain, and transient currents, the architecture drives rapid subspace rotations (behavior switching) and fine-grained trajectory control (within-behavior execution). Validated via a robotic maze simulation where an agent uses sensory feedback to reconfigure its manifold geometry.

Why Neural Manifolds

A neural manifold is the geometric manifestation of a population's progression through a low-dimensional state space. Sequential activity in brain/spinal cord evolves along dynamic, low-dimensional manifolds. Specific circuit mechanisms map manifold geometry to behavior:

• Heterogeneous inhibition → subspace reorientation (rotate sequence into new hyperplanes to switch behavioral states)
• Gain modulation + transient currents → direct control of trajectory velocity and shape

Architecture: Ring Attractor Spiking Network

Implemented as a ring network with asymmetric recurrent connectivity, forming a stable bump of activity that progresses along the ring → oscillatory sequences (Figure 2).

Three Control Neurons

1. Speed control — multiplicative gain S scales how fast the bump travels
2. Shape control — additive current I changes bump width
3. Selection control — subspace inhibition p_inh silences subsets of neurons, steering the trajectory into a target subspace

Control Mechanisms (verified on SpiNNaker 2)

Mechanism	Effect on manifold
Multiplicative gain S ↑	Faster bump traversal (trajectory speed)
Additive current I	Larger bump size during subspace (trajectory shape)
Subspace inhibition p_inh	Rotates activity into a different subspace (behavior switch), sequence dynamics preserved
Combined	Multiple subspaces at increasing rotation counts per unit time

• Subspace rotation verified by first principal angle between subspaces (matches analytical solution)
• Selective inhibition: 80% subspace inhibition leaves sequential dynamics + behavioral state representation intact

SpiNNaker 2 Implementation Details

• Original model is rate-based; SpiNNaker 2 is spike-optimized → added a spike-based communication layer: rate r treated as probability of spiking in current timestep (probabilistic rate→spike)
• Introduced 50% connection sparsity (weights scaled 2× to compensate) to cut incoming spikes per neuron
• Used circulant weight matrix structure → store only one row + 1-bit sparsity mask (memory efficient)
• Host interface streams control parameters in, receives output spikes, computes motor commands
• Runtime scales with: number of neurons, connection sparsity, number of timesteps

Closed-Loop Maze Validation

• Environment: maze from labmaze library; agent moves in real-valued steps, can jump hurdles
• Three subspaces encode three movements: forward, curved-forward, turn
• Agent has a pre-learned world model / plan; the framework translates plan → manifold control parameters (speed, shape, selection)
• Ring-network spikes → readout → motor speeds for two wheels
• Result: agent navigates maze successfully via closed-loop sensory feedback reconfiguring manifold geometry

When to Use This Skill

• Building neuromorphic / SNN controllers for autonomous agents
• Needing explainable neural computation (manifold geometry ↔ behavior mapping)
• Closed-loop control where sensory feedback must reconfigure internal dynamics in real time
• Mapping a high-level plan/world-model onto low-level motor primitives
• Ring-attractor or bump-attractor based sequential generation on constrained hardware

Implementation Checklist

1. Build ring network with asymmetric recurrent weights (circulant + sparse)
2. Implement 3 control inputs: gain (speed), current (shape), inhibition mask (selection)
3. Convert rate→spike probabilistically; keep spike-based inter-neuron communication
4. Assign neuron groups to subspaces via bitmask (1 = in subspace)
5. Decode bump position / spike counts → readout → motor commands
6. Close the loop: environment feedback → update control params → re-steering
7. Verify subspace rotations via principal-angle analysis against analytical solution

Biological & Systems Significance

• Bridges neuroscience (neural manifolds in cortex/spinal cord) and neuromorphic engineering
• Simple circuit mechanisms (gain, inhibition, transient input) readily implement sequence control
• Offers a substrate for investigating biological neural dynamics and for humanoid/robotic control with many DOF
• Framework is general: any plan can be translated into manifold representation suitable for spiking hardware

Pitfalls

• On-chip SRAM is limited: circulant+sparse storage needed; max execution time bounded by local memory
• Higher connectivity → higher runtime per timestep on SpiNNaker 2 (benchmark before scaling)
• Plan was manually created per maze in the paper — learning the plan from observations is future work
• Ring network is powerful but simplified; real networks use dendrites in 3D (extension noted)
• Rate→spike conversion introduces sampling noise; calibrate spike probability to match rate