PaperlayerView exact revision on arXiv

Source-linked AI summary

DCRA: Diffusion-Conditioned Representation Alignment for Robust Time-Series Learning

Wenrui Xu, Anas Enanaa, Keshab K. Parhi

arXiv:2609.11997v1cs.LGstat.ML

TL;DR

Time-series representations remain vulnerable to noise and distribution shifts, a challenge for clinical signal analysis. DCRA uses forward diffusion as a structured corruption schedule and aligns features across noise levels. On CHB-MIT, it improves robustness under multiple noise conditions and sensitivity at low false-positive rates.

  • Problem

    Time-series encoders can degrade under low-SNR conditions and distribution shifts, while common augmentation uses independent perturbations without enforcing representation invariance.

    Source: Introduction · p. 1

  • Method

    DCRA uses forward diffusion as a structured corruption scheduler and aligns clean and corrupted signals through feature-level consistency.

    Source: Introduction · p. 1

  • Results

    DCRA improves robustness across multiple noise levels and sensitivity at low false-positive rates on CHB-MIT.

    Source: Introduction · p. 2, Conclusion · p. 5

  • Takeaways & Limitations

    Structured corruption combined with representation alignment is an effective strategy for robust time-series representation learning.

    Source: Conclusion · p. 5

Abstract

from arXiv · show

Learning robust representations for time-series signals under noise and distribution shifts remains challenging, especially in clinical applications such as electroencephalogram (EEG) and electrocardiogram (ECG) analysis. We propose Diffusion-Conditioned Representation Alignment (DCRA), a training framework that repurposes the forward diffusion process as a structured corruption scheduler for representation learning. Different from conventional augmentation and consistency-based methods that rely on independently sampled perturbations, DCRA introduces a structured corruption trajectory via the diffusion forward process, which enables continuous and controlled representation evolution across noise levels. We introduce a feature-level consistency objective that aligns representations across noise levels while preserving class-discriminative structure. This mechanism promotes structure-preserving consistency, which enables smooth and semantically coherent feature trajectories in latent space. The proposed framework is encoder-agnostic and can be integrated with state space models and Transformer architectures. The seizure detection experiments on the CHB-MIT EEG dataset show that DCRA consistently improves performance under multiple noise conditions and achieves higher sensitivity at low false-positive rates. Analysis reveals that DCRA produces more balanced and structured representations compared to baseline and diffusion-only models. These findings highlight the benefit of combining structured corruption with representation alignment for robust time-series learning.

Introduction

Time-series encoders can degrade under low-SNR conditions and distribution shifts, while common augmentation treats noise as independent perturbations. DCRA uses forward diffusion corruption and feature-level alignment to learn robust, class-discriminative representations.

Introduction

  • Modern sequence encoders can degrade under low-SNR conditions or distribution shifts, while seizure detection requires sensitivity under strict false-positive-rate constraints.The paper frames robustness as important for clinical EEG and ECG applications.Source: Introduction · p. 1
  • Conventional augmentation often treats noise as independent perturbations without enforcing representation invariance across corruption levels.The paper motivates structured corruption trajectories as an alternative to overfitting specific corruption patterns.Source: Introduction · p. 1
  • DCRA combines forward diffusion as a structured corruption scheduler with feature-level consistency between clean and corrupted signals.The encoder receives multiple corruptions of the same input, with alignment encouraging semantically consistent, class-discriminative latent features.Source: Introduction · pp. 1–2
  • On CHB-MIT, DCRA improves robustness across multiple noise levels and sensitivity at low false-positive rates.The paper evaluates ROC-AUC, PR-AUC, and sensitivity under fixed FPR.Source: Introduction · pp. 1–2

Related Work

Prior augmentation and consistency methods generally do not model corruption as a structured trajectory across noise levels. DCRA uses forward diffusion corruption and cross-noise representation alignment as an alternative for robust physiological-signal learning.

Related Work

  • Many augmentation and consistency approaches treat perturbations as independent transformations rather than structured corruption trajectories.DCRA instead generates degraded signals at different noise levels and aligns representations along the diffusion path.Source: Representation Learning via Data… · p. 2

Diffusion Model for Structured Corruption

  • Prior diffusion-based representation approaches often rely on reverse denoising or reconstruction, whereas DCRA uses forward diffusion as a structured noise scheduler within supervised learning.This constrains representation evolution across noise levels.Source: Diffusion Model for Structured… · p. 2

Seizure Detection

  • Physiological-signal robustness is clinically important, but existing EEG and ECG work rarely enforces representation consistency across multiple corruption levels.The framework addresses this through structured diffusion conditioning and cross-noise regularization.Source: Seizure Detection · p. 2

Background

The background describes state-space sequence modeling and diffusion-based progressive corruption. DCRA uses only the forward diffusion process as a structured scheduler for learning representations stable across signal-to-noise ratios.

State Space Models and Mamba

  • The continuous state-space system is discretized with the Zero-Order Hold rule for discrete, high-frequency signals.The state transition, input, and output matrices are denoted A, B, and C, respectively.Source: State Space Models and Mamba · p. 2
  • Mamba adds input-dependent state updates to standard state-space models and has linear complexity with sequence length while retaining long-range modeling capability.This makes it suitable for long, high-frequency time-series analysis.Source: State Space Models and Mamba · pp. 2–3
  • Diffusion probabilistic models progressively add Gaussian noise over timesteps, with the forward-process noise schedule controlling injected-noise magnitude.DCRA uses this forward process, rather than reverse denoising, as a structured corruption scheduler across varying SNRs.Source: Diffusion Probabilistic Models and… · p. 3

Methodology

DCRA generates two corruption levels from each signal and processes them with a shared timestep-conditioned encoder. It combines supervised classification with consistency objectives to align representations across noise levels while retaining class structure.

Dual-timestep Diffusion Engine

  • DCRA creates two views of each input through forward diffusion, using separate low- and high-noise timestep ranges for controlled corruption.The threshold separating the ranges is dynamically selected.Source: Dual-timestep Diffusion Engine · p. 3

Diffusion-conditioned Encoder

  • The shared encoder receives clean and corrupted signals and uses a normalized diffusion timestep, mapped through an MLP, to condition its features on corruption intensity.A bidirectional Mamba encoder is used to model long-range, non-causal biological-signal dependencies; the framework is not restricted to that architecture.Source: Diffusion-conditioned Encoder · p. 3

Joint Optimization Objective

  • Supervised classification uses focal losses for predictions at two noise levels, with α and β controlling their respective contributions.Focal loss is adopted to address class imbalance in biomedical datasets.Source: Joint Optimization Objective · p. 3
  • Cross-noise consistency aligns noisy features toward clean-feature anchors and uses learnable class prototypes to preserve discriminative structure.The consistency objective promotes smooth changes across corruption levels while preserving class-dependent latent geometry.Source: Joint Optimization Objective · p. 4
  • The final objective combines classification and consistency losses, whose components encourage coherent feature evolution and class separation across noise levels.The consistency loss includes anchor-based subspace consistency and prototype-based alignment.Source: Joint Optimization Objective · pp. 3–4

Experimental Setup

The evaluation uses patient-specific seizure detection on CHB-MIT, with seizure-event folds designed to test generalization to unseen events. Preprocessing and normalization are controlled to avoid leakage, and the setup compares baseline encoders with diffusion-conditioned variants.

Experimental Setup

  • Each patient’s model is trained and tested independently, with folds holding out at least one seizure segment from training and validation to assess performance on unseen seizure patterns.The protocol uses multifold cross-validation.Source: Experimental Setup · p. 4
  • EEG recordings are preprocessed into 18 standard channels and segmented into non-overlapping 2-second windows.Source: Experimental Setup · p. 4
  • Training-set z-score normalization is applied to validation and test data, and baseline encoders are compared with diffusion-conditioned versions with and without feature consistency.The controlled comparison is intended to attribute performance differences to DCRA.Source: Experimental Setup · pp. 4–5

Experimental Results & Analysis

Across subjects, DCRA improves seizure-detection sensitivity under strict false-positive-rate constraints while retaining or improving PR-AUC. Under increasing corruption, it preserves discriminative performance and produces more balanced, coherent feature trajectories.

Experimental setup

  • The evaluation averages results across subjects, uses T = 1000 and Tthres = 500, and applies EEGNet as a baseline.The analysis examines performance, robustness, and latent-space representation behavior.Source: Experimental Results & Analysis · p. 5

Performance

  • DCRA improves sensitivity under strict FPR constraints while maintaining or improving PR-AUC, despite only modest gains in global AUC.Feature consistency improves sensitivity at low-FPR operating points without degrading overall performance.Source: Experimental Results & Analysis · p. 5

Robustness

  • As noise increases, all encoders degrade, but DC and DCRA outperform pure encoders; DCRA retains higher AUC and PR-AUC at intermediate-to-high noise.DCRA also achieves higher sensitivity at low FPR for t ≤500 (SNR ≥−20 dB), and degrades more slowly at extreme corruption, including t = 800 (SNR ≤−35 dB).Source: Experimental Results & Analysis · p. 5

Representation analysis

  • Transformer positive-label class cosine similarity rises from 0.898 to 0.945, while Mamba rises from 0.862 to 0.937 with DCRA.Baseline models have the highest raw cross-timestep cosine similarity but show imbalanced intra-class consistency; DCRA maintains comparable similarity with better class balance.Source: Experimental Results & Analysis · p. 5
  • Baseline features evolve rigidly, while DC features follow irregular paths; DCRA produces smoother trajectories aligned with the data geometry.DCRA also avoids DC's excessive feature drift, yielding shorter, more coherent trajectories.Source: Experimental Results & Analysis · p. 5

Interpretation

  • Diffusion time embedding supports robustness to noise, while feature consistency shapes representations into a more semantically meaningful structure.The resulting features are described as more stable and interpretable, with potential usefulness for downstream tasks.Source: Experimental Results & Analysis · p. 5

Conclusion

DCRA uses forward diffusion as structured corruption and cross-noise consistency to learn robust time-series representations. On CHB-MIT, it improves robustness under low-SNR and low-FPR conditions and yields more stable, semantically coherent features.

Conclusion

  • DCRA improves robustness under low-SNR and low-FPR conditions on the CHB-MIT dataset.The framework repurposes the forward diffusion process as a structured corruption scheduler for time-series encoders.Source: Conclusion · p. 5
  • Cross-noise consistency regularization encourages smooth, class-discriminative feature evolution and more coherent representations than baseline and diffusion-conditioned-only models.The paper presents structured corruption combined with representation alignment as an effective strategy for robust time-series representation learning.Source: Conclusion · p. 5
  • Future work will extend DCRA to broader encoder architectures and investigate other time-series learning domains.Source: Conclusion · p. 5
  • Figure 2 reports stronger robustness for DCRA than baseline models as diffusion steps increase.The figure compares different encoders and strategies across timesteps.Source: Figure 2 · p. 6
  • Figure 3 shows smoother DCRA feature trajectories and clearer class-dependent structure as noise increases.Its columns show sample-wise trajectories, domain-shift behavior, and class-colored feature distributions; rows compare baseline, DC, and DCRA encoders for each backbone.Source: Figure 3 · p. 7

Related papers

Loading 2609.11997v1…