| name | dynamic-neural-manifolds-snn-control |
| description | Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware. Spiking ring networks with control knobs (gain, inhibition, transient currents) that steer low-dimensional manifold geometry for explainable autonomous behavior. Implemented on SpiNNaker 2 chip. Activation: neural manifolds, neuromorphic control, SpiNNaker, ring network, subspace rotation, explainable SNN, closed-loop control, dynamic manifolds. |
| tags | ["neuroscience","neuromorphic","SNN","control","manifolds","SpiNNaker","explainable-AI"] |
| version | 1.0.0 |
| author | agent |
| date | 2026-07-12T00:00:00.000Z |
Dynamic Neural Manifolds for Flexible Closed-Loop Control on Neuromorphic Hardware
arXiv: 2607.07373 | cs.NE | von Seeler, Tetzlaff, Lehr (Göttingen)
Overview
Biological circuits evolve sequential neural activity along dynamic, low-dimensional manifolds to enable flexible behavior. This paper presents a framework for implementing dynamic neural manifolds on the SpiNNaker 2 neuromorphic chip for real-time, closed-loop control, establishing an explainable approach to neuromorphic engineering.
Core Methodology
Ring Network Architecture
- Canonical oscillatory sequences modeled as a bump of activity moving along a ring of neurons
- Asymmetric recurrent connectivity leads to stable bump progression → oscillatory sequences
- Circulant weight matrix with configurable sparsity (e.g., 50%) for efficiency
- Maps to observed oscillatory sequences across brain regions (spinal cord, MEC, motor cortex)
Three Control Knobs
-
Trajectory Shape (Additive current I): Controls bump size → trajectory radius in state space
- Excitatory current (I>0) increases bump size and radius
- Maps to spatial extent of neural representation
-
Trajectory Speed (Multiplicative gain S): Controls how fast bump travels around ring
- Gain amplifies inputs → faster sequence progression
- Maps to velocity/timing of neural trajectories
-
Subspace Rotation (Heterogeneous inhibition p_inh): Selective silencing of neuron subsets
- Inhibitory ensembles silence random subsets of neurons
- Switching ensembles rotates neural subspace → enables behavior switching
- Angle between orientations: arccos(1 - p_inh)
- Each subspace encodes a different behavior/motor output
SpiNNaker 2 Implementation
- Spike-based communication: Rates → probabilistic spikes for neuron-to-neuron communication
- Sparse connectivity: Reduces computation (e.g., 20-50% sparsity with weight scaling 1/p)
- Circulant structure: Store single row + bitmask → saves memory
- Streaming control: Control parameters streamed in during runtime (no pre-specified sequences)
- Closed-loop latency: As low as 3 time steps (~3ms) from spike to control update
Key Equations
x_i(t+1) = Σ_j W_ji * r_j(t) + I(t)
r_i(t+1) = F(r_i(t) + 1/τ * (-r_i(t) + p_i(t) * S(t) * x_i(t+1)))
where W is circulant, p_i indicates active (non-inhibited) neurons, S is speed control, I is shape control.
Motor Readout
- Linear readout weights learned from ring network spike activity
- Spike smoothing via exponential moving average: s̄_i(t) = (1-α)s_i(t) + α·s̄_i(t-1)
- Motor speeds and actions decoded as: y_x = (s̄_1, ..., s̄_n, 1) · w_x
Applications
Robotic Navigation
- 2-wheeled robot navigating maze with sensory feedback
- Three subspaces: move forward, turn in place, jump
- Sensory inputs modulate control parameters → dynamic manifold reconfiguration
- High-level plan + local adaptation → integrated manifold representation
Biological Research
- Computational testbed for investigating biological neural dynamics
- Validates theoretical predictions from Lehr et al. (2024, 2025) on subspace rotations
- Bridge between circuit mechanisms and geometric manifold properties
Key Findings
- Dynamic manifold control works robustly on neuromorphic hardware
- Runtime scales linearly with spike count → efficient for sparse activity
- All three control mechanisms (shape, speed, rotation) reproduce theoretical predictions
- Subspace rotations enable behavioral switching with predictable geometry
- Framework provides mathematically interpretable internal states → explainable AI
Pitfalls
- SRAM memory constraint: Recording rates fills 128KB SRAM quickly; use streaming instead
- Memory limits simulation time: Streaming control parameters avoids pre-specified sequences
- Weight precision: 8-bit integers with shared exponent; precision loss for large networks
- Latency budget: Must complete all updates within 1ms time step on each PE
Related Skills
snn-learning-survey - SNN learning rules
spiking-neural-network-analysis - SNN paper analysis
spiking-oscillation-mapping - Oscillatory states in balanced SNNs