Daily Arxiv

This page organizes papers related to artificial intelligence published around the world.
This page is summarized using Google Gemini and is operated on a non-profit basis.
The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.

Diffusion-Augmented Contrastive Learning: A Noise-Robust Encoder for Biosignal Representations

Created by
  • Haebom

Author

Rami Zewail

DACL: Diffusion-Augmented Contrastive Learning for Biosignals

Outline

This paper proposes Diffusion-Augmented Contrastive Learning (DACL), a novel hybrid framework that combines diffusion models and supervised contrastive learning for robust representation learning of biosignals. This framework operates on the latent space generated by a lightweight Variational Autoencoder (VAE) trained on Scattering Transformer (ST) features. The forward process of the diffusion model is utilized as a data augmentation technique to generate multiple noisy views of the latent embedding, and a U-Net-style encoder is used to learn noise-robust and class-separable representations at various diffusion time steps using a supervised contrastive learning objective. Experimental results on the PhysioNet 2017 ECG dataset demonstrate an area under the curve (AUROC) of 0.7815.

Takeaways, Limitations

•
Takeaways:
◦
We propose the possibility of improving the performance of representation learning of biosignals by utilizing diffusion models for data augmentation.
◦
We demonstrate the feasibility of learning noise-robust and class-discriminative embeddings using supervised contrastive learning objectives.
◦
We achieved competitive AUROC on the PhysioNet 2017 ECG dataset.
•
Limitations:
◦
Further research is needed to evaluate the generalization performance of the proposed method and its applicability to other biosignal datasets.
◦
Additional research on computational complexity and model tuning may be required.
◦
Evaluation was performed only on a single dataset (PhysioNet 2017 ECG).
👍
Made with Slashpage