Abstract
Self-supervised learning (SSL) is a powerful paradigm for learning from unlabeled time-series data. However, popular methods such as masked autoencoders (MAEs) rely on reconstructing inputs from a fixed, predetermined masking ratio. Instead of this static design, we propose treating the corruption level as a new degree of freedom for representation learning, enhancing flexibility and performance. To achieve this, we introduce the Flow-Guided Neural Operator (FGNO), a novel framework combining operator learning with flow matching for SSL training. FGNO learns mappings in functional spaces by using Short-Time Fourier Transform to unify different time resolutions. We extract a rich hierarchy of features by tapping into different network layers and flow times that apply varying strengths of noise to the input data. This enables the extraction of versatile representations, from low-level patterns to high-level global features, using a single model adaptable to specific tasks. Unlike prior generative SSL methods that use noisy inputs during inference, we propose using clean inputs for representation extraction while learning representations with noise; this eliminates randomness and boosts accuracy. We evaluate FGNO across three biomedical domains, where it consistently outperforms established baselines. Our method yields up to 35% AUROC gains in neural signal decoding (BrainTreeBank), 16% RMSE reductions in skin temperature prediction (DREAMT), and over 20% improvement in accuracy and macro-F1 on SleepEDF under low-data regimes. These results highlight FGNO's robustness to data scarcity and its superior capacity to learn expressive representations for diverse time series.
Introduction
Self-supervised learning (SSL) has emerged as a critical tool for time-series analysis, particularly in biomedical fields where labeled data is scarce. However, existing methods often struggle with rigid representations derived from fixed masking ratios, resolution mismatches across different sampling rates, and the introduction of noise during inference.
To address these challenges, we introduce the Flow-Guided Neural Operator (FGNO). FGNO treats the generative noise level—referred to as "flow time"—as a tunable parameter. By operating in a resolution-invariant function space, FGNO extracts multi-scale representations from a single pretrained model, achieving state-of-the-art performance and remarkable data efficiency on diverse biomedical benchmarks.
Methods
FGNO integrates operator learning with flow matching to create a flexible SSL framework. The process begins by embedding raw 1D signals into 2D time-frequency spectrograms using the Short-Time Fourier Transform (STFT). This transformation embeds data into a continuous function space, making the model inherently invariant to the original signal's sampling resolution.
During pre-training, the model (a Transformer) is trained to simulate a continuous normalizing flow, learning the vector field that maps Gaussian noise to clean spectrograms. This process is conditioned on a flow time $s \in [0, 1]$, which dictates the noise level. For downstream tasks, we freeze the backbone and extract features using clean inference. Instead of adding actual noise, we feed clean spectrograms to the model while conditioning it on a specific flow time $s$. This allows for deterministic, stable feature extraction across a 2D grid of network layers $l$ and flow times $s$.
Experimental Setup
We evaluated FGNO across several diverse biomedical benchmarks:
- DREAMT: Smartwatch and clinical PSG data for sleep/wake classification and skin temperature regression.
- BrainTreeBank: Intracranial neural responses for speech presence, volume, and pitch classification.
- Epileptic Seizure Recognition: EEG recordings for binary seizure detection.
- SleepEDF: Whole-night EEG for 5-class sleep stage classification.
Results and Discussion
FGNO demonstrates that flow time acts as a semantic dial for feature granularity. High corruption (low $s$) encourages the capture of abstract, global features suitable for regression, while low corruption (high $s$) preserves the fine temporal details necessary for classification.
Key results include:
- Superior Accuracy: FGNO achieved a 35% AUROC improvement on BrainTreeBank and a 16% RMSE reduction on DREAMT compared to strong baselines like MAE and Chronos.
- Data Efficiency: With only 5% of labeled data, FGNO matched the performance of baselines using 100% of labels on SleepEDF (~93.5% accuracy).
- Resolution Invariance: When inputs were downsampled by 48x, FGNO maintained >74% AUROC, while the MAE baseline dropped to ~52%.
- Efficiency: FGNO utilizes only 370K parameters, making it significantly smaller and faster to fine-tune than models like BrainBERT (43M parameters).
Conclusion
FGNO offers a robust and efficient framework for time-series SSL by leveraging the flexibility of flow matching and the resolution invariance of neural operators. While current implementation requires a manual grid search to find the optimal $(l, s)$ pair for new tasks, future work will focus on automating this selection process and expanding the framework to non-biomedical modalities.
Summary updated