Simple SSM MS4 outperforms complex Mamba models on TSC benchmarks
New lightweight MS4N matches 10x larger models on 59 datasets, 50K timesteps
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Researchers at Monash University (Saadatmand, Webb, Rezatofighi, Salehi) present the first systematic comparison of diagonal SSMs (S4D) and input-dependent SSMs (Mamba family) for multivariate time series classification (TSC). Their surprising finding: the simpler S4D consistently beats Mamba-based models in both accuracy and efficiency across large-scale benchmarks. Building on this, they introduce MS4, a lightweight modification of S4D with a linear input projection and channel-mixing mechanism, and MS4N, a normalized version that stabilizes state dynamics with minimal overhead.
Evaluated on 59 datasets from MONSTER (up to 60M samples, 50K timesteps, 82 classes) and UEA, against 15 baselines, MS4 and MS4N outperform all Mamba variants while remaining more efficient. MS4N matches or surpasses competing deep learning models that are roughly 2x and 10x larger in parameters. This positions lightweight structured SSMs as a compelling alternative to scaling complexity for TSC, with implications for resource-constrained deployment in IoT, healthcare, and finance.
- S4D diagonal SSMs consistently outperform Mamba-style input-dependent SSMs on TSC accuracy and efficiency
- MS4N achieves competitive results against models 2-10x larger while using fewer parameters
- Evaluation across 59 datasets, up to 60M samples, 50K timesteps, and 82 classes
Why It Matters
Smarter, simpler models can replace bloated architectures for time-series AI in real-world applications.