Research map / Time series / Causal and generative time series
SOI: Scaling Down Computational Complexity by Estimating Partial States of the Model
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 Scattered Online Inference (SOI), which compresses time-series data with strided convolutions and extrapolates (duplicates) missing partial network states so that the model skips recomputation on alternating inferences, cutting compute for speech separation and acoustic scene classification.
Problem
Neural networks are too computationally expensive for real-time, single-frame online inference on resource-constrained microcontrollers, and prior STMC-style incremental inference either keeps full accuracy at exponentially growing memory cost or needs routing that is not structured for strided laye
Approach
Introduce strided-cloned convolution pairs (strided convolution for compression plus frame duplication as extrapolation) and optional time shifts into a causal U-Net/CNN so that deeper layers are skipped on predetermined inferences, with partially predictive and fully predictive inference patterns and skip connections to restore current-frame information.
Key results
- In speech separation with FP SOI, 50% complexity reduction at a cost of 11.3% SI-SNRi relative to STMC (S-CC 1|3 reaches 6.82 dB vs 7.69 dB for STMC).
- In acoustic scene classification, SOI matched or exceeded STMC accuracy for several models (e.g., model VII 83.35% vs 83.07%) while reducing complexity by roughly 16% versus STMC.
- Compared with resampling baselines at ~909.6 MMAC/s, S-CC 5 reaches 7.47 dB SI-SNRi at 1178.7 MMAC/s versus 3.49-5.83 dB for resampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply PP SOI with one S-CC pair at layer 5 instead of STMC | U-Net · DNS speech separation | Complexity (MMAC/s): 1819.2 -> 1178.7 | Helped |
| Apply PP SOI with S-CC pairs at layers 1 and 3 instead of STMC | U-Net · DNS speech separation | SI-SNRi (dB): 7.69 -> 6.27 | Hurt |
| Apply PP SOI with S-CC pairs at layers 1 and 3 instead of STMC | U-Net · DNS speech separation | Complexity (MMAC/s): 1819.2 -> 528.8 | Helped |
| Replace ordinary convolutions with strided convolutions in predictive inference (ablation) | U-Net predictive inference · DNS speech separation, prediction length 3 | SI-SNRi (dB): 4.61 -> 5.47 | Helped |
| Replace ordinary convolutions with strided convolutions in predictive inference (ablation) | U-Net predictive inference · DNS speech separation, prediction length 1 | SI-SNRi (dB): 7.41 -> 7.24 | Hurt |
| Replace frame duplication with bilinear interpolation (ablation) | PP SOI, S-CC 5 · DNS speech separation | SI-SNRi (dB): 7.47 -> 7.63 | Helped |
| Replace frame duplication with transposed convolution (ablation) | PP SOI, S-CC 1|3 · DNS speech separation | SI-SNRi (dB): 6.27 -> 6.30 | No clear effect |
| Apply SOI rather than STMC | GhostNet I · TAU Urban Acoustic Scene 2020 Mobile | Top-1 accuracy (%): 55.68 -> 55.90 | No clear effect |
| Apply SOI rather than STMC | GhostNet II · TAU Urban Acoustic Scene 2020 Mobile | Top-1 accuracy (%): 64.18 -> 61.98 | Hurt |
| Apply SOI rather than STMC | GhostNet III · TAU Urban Acoustic Scene 2020 Mobile | Top-1 accuracy (%): 66.45 -> 68.14 | Helped |
| Apply SOI rather than STMC | GhostNet VII · TAU Urban Acoustic Scene 2020 Mobile | Top-1 accuracy (%): 83.07 -> 83.35 | No clear effect |
| Use SOI instead of STMC | GhostNet VII · TAU Urban Acoustic Scene 2020 Mobile | Complexity (MMAC/s): 20.87 -> 17.59 | Helped |
Surprising findings
- Strided convolutions gave better predictive performance than standard convolutions for longer prediction horizons (e.g., 4-frame prediction: 4.00 dB vs 3.59 dB).
- Interpolation (bilinear, bicubic) gave slightly better results than frame duplication but adds latency of one frame.
Limitations
- Performance drop in speech separation makes SOI unsuitable where highest accuracy is required.
- Configuration choice (position of S-CC/SC layers) requires tuning and validation; cumulative errors may appear over long sequences.
- Demonstrated only on U-Net and GhostNet-style CNNs and two audio tasks.
Experimental setup
Datasets: DNS Challenge Interspeech 2020 TAU Urban Acoustic Scene 2020 Mobile
Benchmarks: DNS Challenge Interspeech 2020 TAU Urban Acoustic Scene 2020 Mobile
Metrics: SI-SNRi (dB) MMAC/s Top-1 accuracy (%) average inference time (ms) peak memory footprint (MB)
Baselines: STMC Baseline non-causal model resampling (SoX, polyphase, Kaiser, linear) unstructured magnitude pruning
Builds on: STMC U-Net GhostNet
Closest papers by meaning
- TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS (ICLR 2026 · similarity 0.82)
- Are Conventional SNNs Really Efficient? A Perspective from Network Quantization (CVPR 2024 · similarity 0.81)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.81)
- Gated Inference Network: Inference and Learning State-Space Models (NeurIPS 2024 · similarity 0.81)
- AtomNet: Designing Tiny Models from Operators Under Extreme MCU Constraints (AAAI 2025 · similarity 0.81)
- Auto-reconfiguration for Latency Minimization in CPU-based DNN Serving (ICML 2025 · similarity 0.81)
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.