Looped SSMs Outperform Standard Models in Time Series Classification
A recent investigation published on arXiv (2605.16048) reveals that looped State Space Models (SSMs) utilizing depth-recurrence—where the same block is reused across multiple layers—either match or exceed the performance of conventional SSMs with distinct parameters. Evaluated on four architectures (LRU, S5, LinOSS, LrcSSM) and six benchmarks for time series classification, looped SSMs deliver better outcomes, even while functioning within a more limited hypothesis space. The researchers clarify that since the larger model encompasses the looped model as a specific instance, the observed performance improvement is not due to expressivity. Rather, the sharing of parameters across depth provides a constructive inductive bias that enhances optimization. This study suggests that depth-recurrence operates independently from sequence-recurrence and offers its own advantages.
Key facts
- Looped SSMs reuse the same block repeatedly across layers.
- Tested on four architectures: LRU, S5, LinOSS, LrcSSM.
- Evaluated on six time series classification benchmarks.
- Looped SSMs closely match or outperform standard SSMs with k·L parameters.
- Looped model operates within a strictly smaller hypothesis space.
- Performance gain attributed to parameter sharing as inductive bias.
- Depth-recurrence is orthogonal to sequence-recurrence.
- Paper published on arXiv with ID 2605.16048.
Entities
Institutions
- arXiv