In safety-critical applications, microcontrollers must meet stringent quality and performance standards, including the maximum operating frequency$F_{\max}$. Machine learning models have proven effective in estimating$F_{\max}$by utilizing data from on-chip ring oscillators. Previous research has shown that increasing the number of ring oscillators on board can enable the deployment of simple linear regression models to predict$F_{\max}$. However, the scarcity of labeled data that characterize this context poses a challenge in managing high-dimensional feature spaces; moreover, a very high number of ring oscillators is not desirable due to technological reasons. By modeling$F_{\max}$as a linear combination of the ring oscillators’ values, this paper employs Compressed Sensing theory to build the model and perform feature selection, enhancing model efficiency and interpretability. We explore regularized linear methods with convex/non-convex penalties in microcontroller performance screening, focusing on selecting informative ring oscillators. This permits reducing models’ footprint while retaining high prediction accuracy. Our experiments on two real-world microcontroller products compare Compressed Sensing with two alternative feature selection approaches: filter and wrapped methods. In our experiments, regularized linear models effectively identify relevant ring oscillators, achieving compression rates of up to 32:1, with no substantial loss in prediction metrics.

The design of sparse neural networks, i.e., of networks with a reduced number of parameters, has been attracting increasing research attention in the last few years. The use of sparse models may significantly reduce the computational and storage footprint in the inference phase. In this context, the lottery ticket hypothesis (LTH) constitutes a breakthrough result, that addresses not only the performance of the inference phase, but also of the training phase. It states that it is possible to extract effective sparse subnetworks, called winning tickets, that can be trained in isolation. The development of effective methods to play the lottery, i.e., to find winning tickets, is still an open problem. In this article, we propose a novel class of methods to play the lottery. The key point is the use of concave regularization to promote the sparsity of a relaxed binary mask, which represents the network topology. We theoretically analyze the effectiveness of the proposed method in the convex framework. Then, we propose extended numerical tests on various datasets and architectures, that show that the proposed method can improve the performance of state-of-the-art algorithms.

Deep neural networks (DNNs) are increasingly used in critical applications from healthcare to autonomous driving. However, their predictions were shown to degrade in the presence of transient hardware faults, leading to potentially catastrophic and unpredictable errors. Consequently, several techniques have been proposed to increase the fault tolerance of DNNs by modifying network structures and/or training procedures, thereby reducing the need for costly hardware redundancy. There are, however, design or training choices whose impact on fault propagation has been overlooked in the literature. In particular, self-supervised learning (SSL), as a pretraining technique, was shown to improve the robustness of the learned features, resulting in better performance in downstream tasks. This study investigates the fault tolerance of several SSL techniques on image classification benchmarks, including several related to Earth Observation. Experimental results suggests that SSL pretraining, alone or in combination with fault mitigation techniques, generally improves DNNs' fault tolerance, although the performance gap vary among datasets and SSL techniques.

The recovery of signals with finite-valued components from few linear measurements is a problem with widespread applications and interesting mathematical characteristics. In the compressed sensing framework, tailored methods have been recently proposed to deal with the case of finite-valued sparse signals. In this letter, we focus on binary sparse signals and we propose a novel formulation, based on polynomial optimization. This approach is analyzed and compared to the state-of-the-art binary compressed sensing methods.

Iterative l1reweighting algorithms are very popular in sparse signal recovery and compressed sensing, since in the practice they have been observed to outperform classical 11 methods. Nevertheless, the theoretical analysis of their convergence is a critical point, and generally is limited to the convergence of the functional to a local minimum or to subsequence convergence. In this letter, we propose a new convergence analysis of a Lasso l1reweighting method, based on the observation that the algorithm is an alternated convex search for a biconvex problem. Based on that, we are able to prove the numerical convergence of the sequence of the iterates generated by the algorithm. Furthermore, we propose an alternative iterative soft thresholding procedure, which is faster than the main algorithm.

In image compression, classical block-based separable transforms tend to be inefficient when image blocks contain arbitrarily shaped discontinuities. For this reason, transforms incorporating directional information are an appealing alternative. In this paper, we propose a new approach to this problem, namely, a discrete cosine transform (DCT) that can be steered in any chosen direction. Such transform, called steerable DCT (SDCT), allows to rotate in a flexible way pairs of basis vectors, and enables precise matching of directionality in each image block, achieving improved coding efficiency. The optimal rotation angles for SDCT can be represented as solution of a suitable rate-distortion (RD) problem. We propose iterative methods to search such solution, and we develop a fully fledged image encoder to practically compare our techniques with other competing transforms. Analytical and numerical results prove that SDCT outperforms both DCT and state-of-the-art directional transforms.

In this paper, we propose a novel sensor selection scheme for networks equipped with energy harvesting sensing devices. Ultimately, the goal is to minimize the reconstruction distortion at the fusion center by selecting a reduced (i.e., sparse) yet informative enough subset of sensors. The solution must also fulfill the causality constraints associated to the energy harvesting process. For a classical formulation, the optimization problem turns out to be non-convex. To circumvent that, we promote sparsity directly in the power allocation vector by introducing a log-sum penalty term in the cost function. The problem can be iteratively solved by resorting to majorization-minimization procedure leading to a stationary point of the solution. Numerical results reveal that, by using a log-sum penalty term, the sensor selection scheme outperforms others based on the ℓ1 norm while making an effective use of the harvested energy.

In this paper, we propose a method for estimating the sparsity of a signal from its noisy linear projections without recovering it. The method exploits the property that linear projections acquired using a sparse sensing matrix are distributed according to a mixture distribution whose parameters depend on the signal sparsity. Due to the complexity of the exact mixture model, we introduce an approximate two-component Gaussian mixture model whose parameters can be estimated via expectation-maximization techniques. We demonstrate that the above model is accurate in the large system limit for a proper choice of the sensing matrix sparsifying parameter. Moreover, experimental results demonstrate that the method is robust under different signal-to-noise ratios and outperforms existing sparsity estimation techniques.

Recently, distributed algorithms have been proposed for the recovery of sparse signals in networked systems, e.g. wireless sensor networks. Such algorithms allow large networks to operate autonomously without the need of a fusion center, and are very appealing for smart sensing problems employing low-power devices. They exploit local communications, where each node of the network updates its estimates of the sensed signal also based on the correlated information received from neighboring nodes. In the literature, theoretical results and numerical simulations have been presented to prove convergence of such methods to accurate estimates. Their implementation, however, raises some concerns in terms of power consumption due to iterative inter-node communications, data storage, computation capabilities, global synchronization, and faulty communications. On the other hand, despite these potential issues, practical implementations on real sensor networks have not been demonstrated yet. In this paper we fill this gap and describe a successful implementation of a class of randomized, distributed algorithms on a real low-power wireless sensor network test bed with very scarce computational capabilities. We consider a distributed compressed sensing problem and we show how to cope with the issues mentioned above. Our tests on synthetic and real signals show that distributed compressed sensing can successfully operate in a real-world environment.

In this paper, we consider the problem of RSS-fingerprinting localization in wireless sensor networks. In particular, inspired by the recent advances in sparse approximation and compressive sensing theory, we propose a localization scheme based on the dictionary design of block-sparse signals. We show via numerical simulations and real experiments that the proposed technique outperforms traditional fingerprinting methods.

The ℓ0/ℓ1-regularized least-squares approach is used to deal with linear inverse problems under sparsity constraints, which arise in mathematical and engineering fields. In particular, multiagent models have recently emerged in this context to describe diverse kinds of networked systems, ranging from medical databases to wireless sensor networks. In this paper, we study methods for solving ℓ0/ℓ1-regularized leastsquares problems in such multiagent systems. We propose a novel class of distributed protocols based on iterative thresholding and input driven consensus techniques, which are well-suited to work in-network when the communication to a central processing unit is not allowed. Estimation is performed by the agents themselves, which typically consist of devices with limited computational capabilities. This motivates us to develop low-complexity and low-memory algorithms that are feasible in real applications. Our main result is a rigorous proof of the convergence of these methods in regular networks. We introduce a suitable distributed, regularized, least-squares functional, and we prove that our algorithms reach their minima using results from dynamical systems theory. Furthermore, we propose numerical comparisons with the alternating direction method of multipliers and the distributed subgradient methods, in terms of performance, complexity, and memory usage. We conclude that our techniques are preferable for their good memory-accuracy tradeoff.

In this paper, we address the problem of distributed support detection of multiple sparse signals with common support. Specifically, signals are acquired by the individual nodes of a network according to the so-called Joint Sparsity Model 2 (JSM-2). By leveraging on this model, we propose a distributed scheme for in-network signal recovery, i.e. not requiring data gathering and processing at a fusion center, based on distributed iterative thresholding and consensus strategies. For the proposed scheme, whose convergence properties we rigorously prove, no a priori knowledge on the non-zero number of entries in the signal vector is required.

In this paper, we present a new recovery algorithm for innetwork compressed sensing from measurements acquired in multi-agent systems. Each agent has to recover a common signal taking advantage of local communication and simple computations. Such distributed problem typically incurs a high energy cost due to inter-node communications. In this paper we propose an iterative distributed algorithm to address this problem, featuring pairwise gossip communications and updates. We propose some theoretical results on its dynamics and numerical comparisons with the most recent approaches proposed in literature. The performance turns out to be competitive in terms of reconstruction accuracy, complexity, and energy consumption required for convergence.

In this paper, we address the problem of distributed sparse recovery of signals acquired via compressed measurements in a sensor network. We propose a new class of distributed algorithms to solve Lasso regression problems, when the communication to a fusion center is not possible, e.g., due to communication cost or privacy reasons. More precisely, we introduce a distributed iterative soft thresholding algorithm (DISTA) that consists of three steps: an averaging step, a gradient step, and a soft thresholding operation. We prove the convergence of DISTA in networks represented by regular graphs, and we compare it with existing methods in terms of performance, memory, and complexity.

We present PISTA, a GPU-accelerated Iterative Soft Thresholding (IST) algorithm for sparse image recovery in Compressive Sensing applications. As the time required to recover an image increases with the number of pixels, GPU-acceleration enables to recover even large images in reasonable time. With respect to equivalent methods, IST-like algorithms have lower computational complexity per-iteration and lower memory requirements, plus the operations are inherently suitable for parallelization. Our experiments show that our algorithm enables a significant reduction in the time required to recover an image even over a highly-optimized CPU-only reference.