dynamic-neural-manifolds-neuromorphic-control

name	dynamic-neural-manifolds-neuromorphic-control
description	Dynamic neural manifolds methodology for flexible closed-loop control on neuromorphic hardware. Uses ring attractor networks with sensory-modulated control neurons (speed, shape, selection) to drive subspace rotations and fine-grained trajectory control in neural state space. Implemented on SpiNNaker 2 chip with robotic maze navigation validation.
version	1.0.0
author	Hermes Agent (from arXiv:2607.07373v1)
license	MIT
metadata	{"hermes":{"tags":["neuromorphic","neural-manifolds","closed-loop-control","spiking-networks","spinnaker2","ring-attractor"],"trigger_words":["dynamic neural manifold","neuromorphic control","spinnaker 2","ring network","closed-loop spiking","subspace rotation","neural trajectory","manifold geometry","sequential neural activity","flexible behavior","activity bump","control neurons"]},"paper":{"arxiv_id":"2607.07373v1","title":"Dynamic neural manifolds for flexible closed-loop control on neuromorphic hardware","authors":["Oskar von Seeler","Christian Tetzlaff","Andrew B. Lehr"],"published":"2026-07-08","categories":["cs.NE"]}}

Dynamic Neural Manifolds for Neuromorphic Closed-Loop Control

Paper

• Title: Dynamic neural manifolds for flexible closed-loop control on neuromorphic hardware
• Authors: Oskar von Seeler, Christian Tetzlaff, Andrew B. Lehr
• Affiliation: University Medical Center Göttingen; Campus Institute Data Science, University of Göttingen; Circulant Labs
• arXiv: 2607.07373v1 [cs.NE] (2026-07-08)
• License: CC BY 4.0

Core Concept

In biological circuits, sequential neural activity evolves along dynamic, low-dimensional manifolds to enable flexible behavior. This paper presents a framework that makes these manifolds parameterizable through specific circuit mechanisms, enabling explainable neuromorphic control on real hardware.

Key Innovation

The architecture maps neural computation onto a geometric manifold framework where:

• Neural population activity = trajectory in N-dimensional state space
• Sequential activity = movement along low-dimensional curved manifold
• Control = modulation of manifold geometry (position, speed, shape) via simple neural mechanisms

Architecture

Ring Attractor Network

Input → Control Neurons → Ring Network → Readout → Motor Output
         (3 types)         (circulant)   neurons    (agent)

Ring Network: Circulant weight matrix with 50% connection sparsity (weights scaled 2x to compensate). Implements a ring attractor that produces a traveling "bump" of activity — a well-studied pattern in neuroscience (head direction cells, spatial navigation).

Three Control Neuron Types

Control Type	Function	Biological Analogy
Speed control	Modulates propagation speed of activity bump around ring	Neuromodulatory gain control
Shape control	Changes width of activity bump (narrow ↔ broad)	Heterogeneous inhibition tuning
Selection control	Selectively inhibits subsets of ring neurons	Targeted inhibitory interneurons

These three simple mechanisms serve as "control knobs" that allow sensory feedback to:

1. Switch behavioral states (e.g., steering ↔ jumping) via subspace rotation