• Type of project: Research Project/Hauptseminar/Thesis (literature review + coding)
  • Contact: han.wang@tu-ilmenau.de

Description:

In model-based deep learning, Algorithm Unrolling turns iterative solvers for inverse problems into trainable neural networks [1]. A cornerstone of this paradigm is Learned ISTA (LISTA), an unrolled proximal-gradient network for sparse recovery that preserves the interpretability and inductive bias of the LASSO while enabling data-driven speed and accuracy gains [2]. This project systematically compares supervised and self-supervised (physics-guided) training for LISTA and its variants [3] [4].

Tasks:

  • Literature review on algorithm unrolling, signal recovery, and LISTA models.
  • Implement deep unrolled neural networks for LISTA variants in the application of multichannel ultrasound imaging.
  • Benchmark supervised and self-supervised LISTA training under different objectives, reporting accuracy, support recovery, and stability.

References:

[1] Monga, Vishal, Yuelong Li, and Yonina C. Eldar. “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing.” IEEE Signal Processing Magazine 38.2 (2021): 18-44.
[2] Wang, Han, et al. “Efficient convolutional forward modeling and sparse coding in multichannel imaging.” 2024 32nd European Signal Processing Conference (EUSIPCO). IEEE, 2024.
[3] Liu, Jialin, and Xiaohan Chen. “ALISTA: Analytic weights are as good as learned weights in LISTA.” International Conference on Learning Representations (ICLR). 2019.
[4] Nagahisa, Koshi, Ryo Hayakawa, and Youji Iiguni. “Comparison between Supervised and Unsupervised Learning in Deep Unfolded Sparse Signal Recovery.” arXiv preprint arXiv:2509.01331 (2025).