Research map / Architectures and efficiency / Spiking networks
BioOSS: A Bio-Inspired Oscillatory State System with Spatio-Temporal Dynamics
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.
Proposes BioOSS, a 2D grid of coupled pressure (p) and velocity (o) neurons governed by discretized wave PDEs with trainable damping and wave speed, evaluated as a sequence model on UEA classification, PPG regression, and long-term forecasting against LRU, S5, and LinOSS.
Problem
Existing oscillatory sequence models capture temporal dynamics but not spatio-temporal wave propagation with spatial structure as in neural circuits.
Approach
A discretized 2D wave-PDE system with interacting pressure-like (p) and velocity-like (o) neurons on a grid, with trainable damping and propagation speed; eigendecomposition in Fourier space yields closed-form complex eigenvalues that enable parallel scan over time.
Key results
- On UEA classification, BioOSS averages 66.25% accuracy vs 65.12% for LinOSS, with best results on Worms (92.8%), EthanolConcentration (33.4%), and Heartbeat (74.8%).
- On PPG-DaLiA, BioOSS reaches MSE 7.7 (x10^-2) vs LinOSS 6.4, placing second.
- On long-term forecasting, BioOSS achieves lower MSE than LinOSS on all four datasets (e.g., Traffic 62.284 vs 68.828).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BioOSS instead of LinOSS | UEA time-series classification · Worms | accuracy (%): 92.2 ± 7.7 -> 92.8 ± 5.2 | Helped |
| Use BioOSS instead of S5 | UEA time-series classification · SCP1 | accuracy (%): 86.8 ± 4.5 -> 85.6 ± 3.9 | Hurt |
| Use BioOSS instead of LinOSS | UEA time-series classification · SCP2 | accuracy (%): 60.4 ± 5.7 -> 55.1 ± 1.8 | Hurt |
| Use BioOSS instead of LinOSS | UEA time-series classification · Ethanol | accuracy (%): 29.9 ± 0.6 -> 33.4 ± 10.7 | Helped |
| Use BioOSS instead of S5 | UEA time-series classification · Heartbeat | accuracy (%): 73.9 ± 3.1 -> 74.8 ± 2.0 | Helped |
| Use BioOSS instead of LinOSS | UEA time-series classification · Motor | accuracy (%): 60.4 ± 7.2 -> 55.8 ± 5.8 | Hurt |
| Use BioOSS instead of LinOSS | PPG heart-rate regression · PPG-DaLiA | MSE × 10^-2: 6.4 ± 0.23 -> 7.7 ± 0.2 | Hurt |
| Use BioOSS instead of LinOSS | multivariate forecasting · Electricity | MSE × 10^-2: 21.875 ± 0.000 -> 18.699 ± 0.000 | Helped |
| Use BioOSS instead of LinOSS | multivariate forecasting · Solar-Energy | MSE × 10^-2: 22.363 ± 0.000 -> 21.873 ± 0.003 | Helped |
| Use BioOSS instead of LinOSS | multivariate forecasting · Traffic | MSE × 10^-2: 68.828 ± 0.002 -> 62.284 ± 0.007 | Helped |
| Use BioOSS instead of LinOSS | multivariate forecasting · Weather | MSE × 10^-2: 22.288 ± 0.000 -> 22.232 ± 0.000 | Helped |
Surprising findings
- LinOSS achieves lower MSE than BioOSS on PPG-DaLiA regression.
- The authors note their reproduced LinOSS numbers differ from those in the original LinOSS paper.
Limitations
- Non-linear dynamics introduce training complexity and hyperparameter sensitivity
- Linear encoder limits expression of intrinsic oscillatory behavior
- Only a partial abstraction of biological signals; no explicit connectivity boundaries
Experimental setup
Datasets: UEA-MTSCA six datasets PPG-DaLiA Electricity Solar-Energy Traffic Weather
Benchmarks: UEA multivariate time series classification (six datasets) PPG-DaLiA heart rate regression Electricity Solar-Energy Traffic Weather
Metrics: accuracy mean-squared error
Baselines: LRU S5 LinOSS
Builds on: LinOSS S5 Linear Recurrent Unit
Closest papers by meaning
- Oscillatory State-Space Models (ICLR 2025 · similarity 0.87)
- From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination (NeurIPS 2026 · similarity 0.87)
- Artificial Kuramoto Oscillatory Neurons (ICLR 2025 · similarity 0.86)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.85)
- TIDE: Asymmetric Neural Circuits for Stabilized Temporal Inhibitory-Excitatory Dynamics (NeurIPS 2026 · similarity 0.85)
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.