LoRD-Net: Unfolded Deep Detection Network With Low-Resolution Receivers

Khobahi, Shahin; Shlezinger, Nir; Soltanalian, Mojtaba; Eldar, Yonina C.

doi:10.1109/tsp.2021.3117503

Cited by 35 publications

(13 citation statements)

References 51 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A further possibility is the use of projected gradient descent, 190 which treats the physical weights as continuous but projects them onto the allowed discrete values after every k th iteration.…”

Section: Back-propagation Under Quantization Constraintsmentioning

confidence: 99%

Intelligent meta-imagers: From compressed to learned sensing

Saigre-Tardif

Faqiri

Zhao

et al. 2022

Applied Physics Reviews

View full text Add to dashboard Cite

Computational meta-imagers synergize metamaterial hardware with advanced signal processing approaches such as compressed sensing. Recent advances in artificial intelligence (AI) are gradually reshaping the landscape of meta-imaging. Most recent works use AI for data analysis, but some also use it to program the physical meta-hardware. The role of “intelligence” in the measurement process and its implications for critical metrics like latency are often not immediately clear. Here, we comprehensively review the evolution of computational meta-imaging from the earliest frequency-diverse compressive systems to modern programmable intelligent meta-imagers. We introduce a clear taxonomy in terms of the flow of task-relevant information that has direct links to information theory: compressive meta-imagers indiscriminately acquire all scene information in a task-agnostic measurement process that aims at a near-isometric embedding; intelligent meta-imagers highlight task-relevant information in a task-aware measurement process that is purposefully non-isometric. The measurement process of intelligent meta-imagers is, thus, simultaneously an analog wave processor that implements a first task-specific inference step “over-the-air.” We provide explicit design tutorials for the integration of programmable meta-atoms as trainable physical weights into an intelligent end-to-end sensing pipeline. This merging of the physical world of metamaterial engineering and the digital world of AI enables the remarkable latency gains of intelligent meta-imagers. We further outline emerging opportunities for cognitive meta-imagers with reverberation-enhanced resolution, and we point out how the meta-imaging community can reap recent advances in the vibrant field of metamaterial wave processors to reach the holy grail of low-energy ultra-fast all-analog intelligent meta-sensors.

show abstract

Section: Back-propagation Under Quantization Constraintsmentioning

confidence: 99%

Intelligent meta-imagers: From compressed to learned sensing

Saigre-Tardif

Faqiri

Zhao

et al. 2022

Applied Physics Reviews

View full text Add to dashboard Cite

show abstract

“…Such unfolded networks depart from the iterative algorithm from which they originates, allowing them to overcome mismatches and approximation errors associated with the need to specify a mathematically tractable surrogate objective for decision making. In particular, training an unfolded network designed with a mismatched model using data corresponding to the true underlying scenario typically yields improved performance compared to the model-based iterative algorithm with the same model-mismatch, as the unfolded network can learn to compensate for this mismatch (30).…”

Section: Example 17mentioning

confidence: 99%

Model-Based Deep Learning: On the Intersection of Deep Learning and Optimization

Shlezinger¹,

Eldar²,

Boyd³

2022

Preprint

Self Cite

View full text Add to dashboard Cite

Decision making algorithms are used in a multitude of different applications. Conventional approaches for designing decision algorithms employ principled and simplified modelling, based on which one can determine decisions via tractable optimization. More recently, deep learning approaches that use highly parametric architectures tuned from data without relying on mathematical models, are becoming increasingly popular. Model-based optimization and data-centric deep learning are often considered to be distinct disciplines. Here, we characterize them as edges of a continuous spectrum varying in specificity and parameterization, and provide a tutorial-style presentation to the methodologies lying in the middle ground of this spectrum, referred to as model-based deep learning. We accompany our presentation with running examples in super-resolution and stochastic control, and show how they are expressed using the provided characterization and specialized in each of the detailed methodologies. The gains of combining model-based optimization and deep learning are demonstrated using experimental results in various applications, ranging from biomedical imaging to digital communications.

show abstract

“…It has been shown empirically that, in comparison with ISTA, LISTA can deliver a more accurate sparse vector with significantly fewer layers/iterations (e.g., see [45]). The success of LISTA has spurred numerous applications of algorithm unrolling over the years (see [89], [112], [69], [9], [118] and the references therein), and, more recently, triggered significant interest in the theoretical foundations of unfolding algorithms.…”

Section: B the Unfolding Principle: Listamentioning

confidence: 99%

Theoretical Perspectives on Deep Learning Methods in Inverse Problems

Scarlett¹,

Heckel²,

Rodrigues³

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

In recent years, there have been significant advances in the use of deep learning methods in inverse problems such as denoising, compressive sensing, inpainting, and super-resolution. While this line of works has predominantly been driven by practical algorithms and experiments, it has also given rise to a variety of intriguing theoretical problems. In this paper, we survey some of the prominent theoretical developments in this line of works, focusing in particular on generative priors, untrained neural network priors, and unfolding algorithms. In addition to summarizing existing results in these topics, we highlight several ongoing challenges and open problems.

show abstract

LoRD-Net: Unfolded Deep Detection Network With Low-Resolution Receivers

Cited by 35 publications

References 51 publications

Intelligent meta-imagers: From compressed to learned sensing

Intelligent meta-imagers: From compressed to learned sensing

Model-Based Deep Learning: On the Intersection of Deep Learning and Optimization

Theoretical Perspectives on Deep Learning Methods in Inverse Problems

Contact Info

Product

Resources

About