Research map / Time series / Foundation models for time series
FIC-TSC: Learning Time Series Classification with Fisher Information Constraint
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 FIC-TSC, a time series classification training framework that constrains the diagonal Fisher information via gradient re-normalization to find flatter minima and improve robustness to train/test domain shift.
Problem
Time series classification suffers from train/test domain shift, and instance normalization (RevIN), effective for forecasting, does not help classification. Standard training also tends to converge to sharp minima that generalize poorly under shift.
Approach
Approximate the Fisher information matrix by its diagonal (squared gradients) and rescale gradients when the entrywise 1-norm exceeds a bound epsilon, so training is a single backward pass while implicitly reducing sharpness; the method is argued to converge at O(1/T).
Key results
- Universal-hyperparameter model reaches 76.9% average accuracy on 26 UEA datasets vs 74.8% for the best prior methods (MILLET, TodyNet), and 86.2% on UCR 85 vs 86.0% for HC2.
- Adding the FIC constraint to ITime, TimesNet, and PatchTST gives around 4% accuracy gains; sharpness is reduced by about 40% on average.
- Compared with SAM on 10 datasets, the method improves accuracy by 2.6% (76.8 vs 74.2) and halves per-iteration runtime (0.070 s vs 0.158 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FIC constraint to InceptionTime (ablation) | InceptionTime · UEA 26 datasets | average accuracy: 72.5 -> 76.9 | Helped |
| Use dataset-wise hyperparameter grid search rather than universal hyperparameters (ablation) | FIC-TSC on UEA 30 datasets · UEA 30 datasets | accuracy: 0.780 -> 0.809 | Helped |
| Use dataset-wise hyperparameter grid search rather than universal hyperparameters (ablation) | FIC-TSC on UCR 85 datasets · UCR 85 datasets | accuracy: 86.2 -> 87.3 | Helped |
| Add FIC constraint (ablation) | InceptionTime · 10 selected UEA datasets | accuracy: +4% average improvement | Helped |
| Add FIC constraint to TimesNet (ablation) | TimesNet · 10 selected UEA datasets | accuracy: around 4% gain in accuracy | Helped |
| Add FIC constraint to PatchTST (ablation) | PatchTST · 10 selected UEA datasets | accuracy: around 4% gain in accuracy | Helped |
| Apply instance normalization (RevIN) (ablation) | Time-series classification models · 10 selected datasets | average accuracy | Hurt |
| Apply instance normalization (RevIN) (ablation) | Time-series classification models · 10 selected datasets | classification performance: p-Value 0.889 | No clear effect |
| Add FIC constraint (ablation) | InceptionTime · TDBrain | accuracy: 93.0 -> 96.2 | Helped |
| Add FIC constraint (ablation) | InceptionTime · ADFTD | accuracy: 46.0 -> 52.8 | Helped |
| Add FIC constraint (ablation) | InceptionTime · PTB-XL | accuracy: 73.9 -> 75.1 | Helped |
| Add FIC constraint (ablation) | InceptionTime · SleepEDF | accuracy: 85.1 -> 86.7 | Helped |
Surprising findings
- Instance normalization (RevIN) yields no significant benefit for classification; Wilcoxon p = 0.889 for with vs without RevIN.
- Within-class train/test distribution distances can equal or exceed between-class distances on some datasets (e.g., FaceDetection, SelfRegulationSCP1).
Limitations
- Empirical convergence may be slower because of the constraint, per the authors' remark.
- Results on 26 UEA datasets only, since several baselines could not run on the remaining datasets; hyperparameter grid search used for the Full setting.
Experimental setup
Datasets: UEA multivariate archive (30 datasets) UCR univariate archive (85 datasets) TDBrain ADFTD PTB-XL SleepEDF
Benchmarks: UEA 30 (26 reported in main table) UCR 85 TDBrain ADFTD PTB-XL SleepEDF
Metrics: accuracy balanced accuracy F1 macro precision recall runtime per iteration sharpness
Baselines: MILLET TodyNet ConvTran ROCKET ITime TsLaNet GPT4TS TimesNet PatchTST CrossFormer TS-TCC HC2
Builds on: ITime RevIN sharpness-aware minimization
Closest papers by meaning
- Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (NeurIPS 2025 · similarity 0.84)
- Con4m: Context-aware Consistency Learning Framework for Segmented Time Series Classification (NeurIPS 2024 · similarity 0.84)
- A Set-Sequence Model for Time Series (NeurIPS 2026 · similarity 0.83)
- T-Rep: Representation Learning for Time Series using Time-Embeddings (ICLR 2024 · similarity 0.83)
- TSLANet: Rethinking Transformers for Time Series Representation Learning (ICML 2024 · similarity 0.82)
- ChronoEpilogi: Scalable Time Series Selection with Multiple Solutions (NeurIPS 2024 · similarity 0.82)
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.