Research map / Time series / Foundation models for time series
Energy Efficient Streaming Time Series Classification with Attentive Power Iteration
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 strAPI, a streaming time series classifier that combines a TCN encoder with an attentive, weighted incremental power-iteration sketch over a supervised latent space, to classify series in real time with low energy use.
Problem
Real-time classification of streaming multivariate time series on resource-constrained devices, where full series are not available and energy, memory, and compute are limited.
Approach
Replaces RNN hidden-state memory with an incrementally updated matrix sketch computed by an attention-weighted batch power iteration (hist-PCA style) over TCN embeddings, trained end-to-end with a classification head, so each update needs only the current batch.
Key results
- Over 70% lower energy consumption than LSTM and 30% (training) / 48% (inference) less than SRNN on average.
- About threefold faster than LSTM and about 1.3x faster than SRNN; MACs roughly one-fifth of LSTM and one-third of SRNN.
- Accuracy gains of 10+% on Epilepsy and 45+% on Handwriting over baselines; comparable on Google-13, DSA-19, HAR-6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use strAPI instead of LSTM | Streaming time series classification · Five benchmark datasets | training energy consumption: over 70% reduction | Helped |
| Use strAPI instead of SRNN | Streaming time series classification · Five benchmark datasets | training energy consumption: 30% less | Helped |
| Use strAPI instead of SRNN | Streaming time series classification · Five benchmark datasets | inference energy consumption: 48% less | Helped |
| Use strAPI instead of LSTM | Streaming time series classification · Five benchmark datasets | running speed: threefold | Helped |
| Use strAPI instead of SRNN | Streaming time series classification · Five benchmark datasets | running speed: 1.3 times faster | Helped |
| Use strAPI instead of LSTM | Streaming time series classification · Five benchmark datasets | MACs: only about one-fifth of LSTM's | Helped |
| Use strAPI instead of SRNN | Streaming time series classification · Five benchmark datasets | MACs: one-third of SRNN's | Helped |
| Use strAPI instead of the baselines | Time series classification · Epilepsy and Handwriting | accuracy: 10+% improvement; 45+% improvement | Helped |
| Remove the Temporal Convolutional Network (ablation) | strAPI · HAR-6 | accuracy | Hurt |
| Replace Attentive Power Iteration (Algorithm 2) with hist-PCA (Algorithm 1) (ablation) | strAPI · HAR-6 | accuracy | Hurt |
| Remove positional encoding (ablation) | strAPI · HAR-6 | accuracy | Hurt |
| Remove embedding projection (ablation) | strAPI · HAR-6 | accuracy | Hurt |
Surprising findings
- On Handwriting and Epilepsy, the gap to baselines widens as series length grows, which the authors attribute to RNN catastrophic forgetting.
Limitations
- Only evaluated on whole-series classification with a single class per series; concept drift not addressed.
- Baselines are limited to three recurrent or transformer models; no comparison with other streaming or sketching methods in the classification setting.
Experimental setup
Datasets: Google-13 DSA-19 HAR-6 Handwriting Epilepsy
Benchmarks: Google-13 DSA-19 HAR-6 Handwriting Epilepsy
Metrics: accuracy energy consumption (kWh) running time (s) multiply-accumulate computations (MACs)
Baselines: LSTM SRNN TRNS (Transformer over SRNN embeddings)
Builds on: hist-PCA Shallow RNN Temporal Convolutional Network Continuous Augmented Positional Embeddings
Closest papers by meaning
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.82)
- Temporal Streaming Batch Principal Component Analysis for Time Series Classification (Student Abstract) (AAAI 2025 · similarity 0.81)
- Spectral-Aware Reservoir Computing for Fast and Accurate Time Series Classification (ICML 2025 · similarity 0.81)
- Amplifier: Bringing Attention to Neglected Low-Energy Components in Time Series Forecasting (AAAI 2025 · similarity 0.81)
- Content-aware Balanced Spectrum Encoding in Masked Modeling for Time Series Classification (AAAI 2025 · similarity 0.81)
- TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation (ICLR 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.