Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data

Duy Nguyen, Jiachen Yao, Jiayun Wang, Julius Berner, Animashree Anandkumar

TL;DR

The Flow-Guided Neural Operator (FGNO) framework redefines self-supervised learning for time-series data by treating noise corruption levels as a continuous degree of freedom rather than a static masking ratio. By combining flow matching with resolution-invariant neural operators, FGNO extracts a rich, task-adjustable hierarchy of features from clean inputs, delivering massive performance gains in data-scarce and multi-resolution biomedical scenarios.

What Problem Does This Paper Address?

Time-series data (such as EEG, ECG, or smartwatch signals) is notoriously messy. It is often recorded at varying sampling rates, and different downstream tasks require analyzing data at completely different scales—some need fine-grained local patterns (spanning seconds), while others require long-term global trends (spanning hours).

Current self-supervised learning (SSL) methods, such as Masked Autoencoders (MAEs), rely on a fixed, static masking ratio during pre-training, yielding a single rigid representation. Meanwhile, generative SSL methods that use diffusion or flow matching typically feed noisy data into the model during inference, introducing unwanted randomness and information loss. The authors tackle these challenges by creating a flexible, resolution-agnostic framework that can extract multi-scale features deterministically.

Methods

FGNO introduces two core innovations to achieve flexible representation learning:

  1. Resolution-Invariant Embedding via STFT: Instead of operating directly on raw 1D time-series vectors, FGNO uses the Short-Time Fourier Transform (STFT) to convert signals into magnitude spectrograms. This functional space representation allows the model to handle varying sampling rates natively without distortion from resampling or interpolation.
  2. Flow Matching with Clean Probing: The model is pre-trained using a simulation-free flow matching objective that learns to map a simple Gaussian noise distribution to the complex data distribution. Crucially, during the downstream "probing" phase, FGNO deviates from standard generative SSL: it feeds clean spectrograms into the frozen backbone. By continuously tuning the conditional flow time ($s$) and selecting different network layers ($l$), users gain a continuous "control knob" over feature granularity. Lower flow times and shallower layers emphasize fine temporal details, while higher flow times and deeper layers capture abstract global semantics.
Overview of FGNO framework compared to standard MAE

Datasets

The authors evaluated FGNO across several challenging biomedical benchmarks:

  • DREAMT: Synchronized smartwatch (Blood Volume Pulse and accelerometer) and clinical polysomnography data used for sleep/wake classification and skin temperature regression.
  • BrainTreeBank: A large-scale dataset of intracranial neural responses from subjects watching movies, evaluated on speech presence classification.
  • Epileptic Seizure Recognition: EEG recordings from 500 subjects used for binary seizure detection.
  • SleepEDF: Whole-night polysomnography EEG data for 5-class sleep stage classification.

Key Insights and Ideas

  • Clean Inputs Beat Noisy Probing: Querying a generative model with clean data and a time-conditioning variable ($s$) yields stable, deterministic features. This completely eliminates the inference variance typically introduced by stochastic noise generation.
  • Task-Dependent Feature Granularity: Different tasks naturally prefer different points in the $(l, s)$ hyperparameter space. For instance, physiological regression tasks rely heavily on global context (lower to intermediate flow times / higher corruption levels), whereas classification tasks requiring precise local timing prefer higher flow times (lower corruption).
  • True Resolution Agnosticism: Because FGNO operates in a functional space defined by STFT, a model trained on high-resolution data generalizes remarkably well to heavily downsampled data, outperforming models that tokenize raw 1D values.

Key Results

  • Biomedical Performance: FGNO achieved up to a 35% AUROC gain in neural signal decoding on BrainTreeBank and a 16% RMSE reduction in skin temperature prediction on DREAMT compared to MAE baselines.
  • Data Efficiency: In low-data regimes (using only 5% of labeled data), FGNO maintained 93.5% accuracy on SleepEDF and 94.1% accuracy on Epilepsy, outperforming strong baselines by over 20% and nearly matching full-dataset performance.
  • Robustness to Downsampling: When tested on BrainTreeBank data downsampled by up to 48x, FGNO maintained an AUROC above 74%, while the MAE baseline dropped sharply to near-random guessing (~52%).
  • Computational Efficiency: By utilizing a lightweight probing head on a frozen backbone, FGNO reduced downstream adaptation time by approximately 60% compared to full-model fine-tuning baselines.

Limitations and Future Work

The primary limitation of FGNO is its current reliance on a discrete grid search to find the optimal layer ($l$) and flow time ($s$) combination for each downstream task. While this grid search is computationally efficient during the probing phase, the authors note that automating this selection process is a key direction for future work. Additionally, they plan to expand the framework's applicability to non-biomedical time-series modalities.

Summary updated

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.
BibTeX
@misc{nguyen2026self,
  author = {Duy Nguyen and Jiachen Yao and Jiayun Wang and Julius Berner and Animashree Anandkumar},
  title = {{Self-Supervised Learning via Flow-Guided Neural Operator on Time-Series Data}},
  year = {2026},
  eprint = {2602.12267},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.12267},
}
Figure