Research map / Probabilistic and scientific modeling / Dynamical systems
Artificial Kuramoto Oscillatory Neurons
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Artificial Kuramoto Oscillatory Neurons (AKOrN), unit-sphere oscillators updated by a generalized Kuramoto dynamics, and shows gains in object discovery, Sudoku reasoning, and adversarial robustness/calibration on CIFAR10.
Problem
Threshold-based neurons lack binding via synchronization, dynamic representations, and built-in reasoning, robustness and calibration; the paper asks whether Kuramoto-style oscillatory neurons can serve as a basic unit of neural networks.
Approach
Replace threshold units with N-dimensional unit-vector oscillators whose iterative updates follow a generalized (vector-valued, asymmetric, symmetry-breaking-input) Kuramoto model, combined with conv or attention connectivity and a norm-based readout that is phase invariant.
Key results
- On PascalVOC, AKOrN achieves MBO_i 52.0 and MBO_c 60.3, outperforming DINO (47.2/53.5) and SPOT (48.3/55.6).
- On Sudoku OOD, AKOrN reaches 89.5±2.5 board accuracy, above IRED and other baselines, and perfectly solves the ID test set.
- Test-time extension of Kuramoto steps raises Sudoku OOD accuracy from 17% to 52%, versus 14% to 34% for standard self-attention steps.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AKOrN instead of existing SSL models | object discovery · PascalVOC | MBO_i: 50.4 -> 52.0 | Helped |
| Use AKOrN instead of SPOT | object discovery · PascalVOC | MBO_c: 55.6 -> 60.3 | Helped |
| Use AKOrN instead of SPOT | object discovery · COCO2017 | MBO_i: 35.0 -> 31.3 | Hurt |
| Use AKOrN instead of SPOT | object discovery · COCO2017 | MBO_c: 44.7 -> 40.3 | Hurt |
| Remove the projection operator (ablation) | AKOrN · CLEVR-Tex | FG-ARI: 80.5\spm{1.5} -> 79.0\spm{2.5} | No clear effect |
| Remove the projection operator (ablation) | AKOrN · Sudoku OOD | accuracy: 51.7\spm{3.3} -> 45.0\spm{1.9} | Hurt |
| Remove the projection operator (ablation) | AKOrN · CIFAR10 | adversarial accuracy: 0.1 -> 64.9 | Hurt |
| Replace the Kuramoto update with a conventional residual update (ablation) | AKOrN · CLEVR-Tex | FG-ARI: 66.2\spm{1.6} -> 80.5\spm{1.5} | Hurt |
| Replace the Kuramoto update with a conventional residual update (ablation) | AKOrN · Sudoku OOD | accuracy: 17.1\spm{16.6} -> 51.7\spm{3.3} | Hurt |
| Reduce rotating dimensions to N=2 (ablation) | AKOrN · object discovery and Sudoku | performance | Hurt |
| Remove the norm-taking term m (ablation) | AKOrN · model fitness | test accuracy: 77.8 -> 84.6 | Hurt |
| Remove conditional stimuli C and use a 3×3 convolution to initialize X (ablation) | AKOrN · model fitness | test accuracy: 62.9 -> 84.6 | Hurt |
Surprising findings
- Adversarially robust to gradient attacks despite training only on clean examples, with strong calibration (near-perfect correlation between confidence and accuracy).
- Models without the symmetric-coupling constraint performed better across tasks than symmetric-J variants, contrary to the Lyapunov-based design intuition.
Limitations
- Hard unit-norm constraint prevents representing feature presence, so poor on memory tasks.
- Performance drops sharply when N exceeds a dataset-dependent threshold; N=2 underfits object discovery and Sudoku.
- Energy is not guaranteed to be a Lyapunov function for attentive connectivity, since connectivity depends on oscillator state.
Experimental setup
Datasets: Tetrominoes dSprites CLEVR Shapes CLEVRTex ImageNet PascalVOC COCO2017 Sudoku CIFAR10 CIFAR10-C
Benchmarks: Tetrominoes dSprites CLEVR Shapes CLEVRTex OOD CLEVRTex CAMO PascalVOC COCO2017 Sudoku ID/OOD CIFAR10 CIFAR10-C
Metrics: FG-ARI MBO board accuracy classification accuracy ECE
Baselines: Slot-attention SLATE DINOSAUR Slot-diffusion SPOT MAE MoCoV3 DINO IRED R-Transformer AutoAttack-evaluated standard CNN/ViT
Builds on: Kuramoto model SimCLR Slot Attention Iterative self-attention (vision transformer) Energy transformer
Closest papers by meaning
- From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination (NeurIPS 2026 · similarity 0.86)
- BioOSS: A Bio-Inspired Oscillatory State System with Spatio-Temporal Dynamics (NeurIPS 2025 · similarity 0.86)
- Winfree Oscillatory Neural Network (NeurIPS 2026 · similarity 0.85)
- Unconditional stability of a recurrent neural circuit implementing divisive normalization (NeurIPS 2024 · similarity 0.84)
- Neurons as Detectors of Coherent Sets in Sensory Dynamics (NeurIPS 2025 · similarity 0.84)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.