This work addresses the linear consensus problem in multi-agent systems under adversarial attacks. We examine scenarios where legitimate agents utilize stochastic inter-agent trust observations to assess the likelihood of neighboring agents acting maliciously in order to mitigate their impact. Malicious agents, in turn, aim to strategically choose what values to send to maximize the disagreement among legitimate agents in a finite horizon. In contrast to prior studies that assume static adversarial behavior and focus on asymptotic consensus, this work investigates the impact of strategic attacks within a finite number of iterations. Specifically, we characterize the maximum disagreement that malicious agents can induce in finite time and the computational complexity of computing the best attack strategy. We compare the effectiveness of different attack strategies through numerical experiments.
We propose a unified multitasking deep learning method to identify and segment structural regions and detect immune cells of non-small cell lung cancer tissue. This approach combines multitask learning and deep supervision to significantly reduce training time compared with single-task models, while achieving competitive performance with 50% fewer parameters. We balance tasks using homoscedastic uncertainty and address class imbalance through a combination of focal loss and class weight adjustment. Tested on multiplexed WSI, our method overcomes challenges related to variability and limited annotations, while extracting robust spatial information from the tumor microenvironment (TME). This data can be leveraged to quantify tumor-immune dynamics and intercellular communication, thereby providing objective insights into the TME's structural and functional complexity.
Deep Reinforcement Learning (DRL) has achieved remarkable success across various domains, yet its real-world applicability remains challenging due to training instability, particularly in complex environments with sparse rewards. Imitation learning mitigates this issue by leveraging expert demonstrations, allowing learning in scenarios where standard RL struggles. Experts are typically humans or teacher agents, often trained with privileged information (PI)-data available during training but not at inference. In this work, we extend beyond standard action replication from teacher to student by also transferring knowledge of latent representations. First, a privileged teacher is trained to use ground-truth information to accelerate learning. Then, a student agent, which lacks access to PI, is trained to align both its actions and the geometry of its intermediate representations with those of the teacher-effectively cloning both decision-making and latent behavior. We evaluate our method on two robotic manipulation tasks and demonstrate that latent behavior cloning significantly improves reward performance and convergence time compared to standard behavior cloning.
Musical instrument classification in the symbolic domain is a challenging task due to the inherent differences between MIDI representations and their corresponding acoustic signals. In this study, we propose a contrastive learning-based framework, Contrastive Audio-MIDI Learning (CAMIL), to improve MIDI-domain instrument classification by leveraging both symbolic and audio information. In our approach, we use MIDI embeddings as anchors, pairing them with their corresponding audio embeddings as positive samples and audio embeddings from different instruments as negative samples. By optimizing a contrastive loss function, our model learns to align MIDI embeddings with their corresponding audio representations while pushing apart embeddings of different instruments. We evaluate our method on the Lakh MIDI dataset and demonstrate that it improves instrument classification performance in the symbolic domain. Our results highlight the potential of contrastive learning in bridging the gap between audio and MIDI representations for more robust musical instrument recognition.
Identification of linear time-varying acoustic systems with multiple inputs and outputs is required in signal processing tasks like echo cancellation or crosstalk cancellation. When all inputs are excited simultaneously, identification is difficult because each output is a superposition of the influence of all inputs. If the inputs are correlated, identification is even more difficult due to the so-called non-uniqueness problem. A recent approach uses an extended Kalman filter to identify acoustic systems on nonlinear lower-dimensional manifolds. We extend this approach to MIMO systems. Instead of simply increasing the size of the neural networks, we propose architectural variants to control the number of parameters. We show that restricting the size of the network in exchange for its flexibility is beneficial for online system identification.
This paper proposes a time-domain approach to analyzing GRF (Ground Reaction Force) signals related to human gait. The method uses the basic half-band FIR filters. It is aimed at extracting features for recognizing the type of shoes worn by a person that trod on a force plate. A basic feature can be the sign of a sample of the first-order difference of the vertical GRF, and higher-level features can be defined by considering series of samesign samples in the forepart of a force wave. The low-frequency component of the difference provides more consistent values of features of the same kind, while in the high-frequency component one can find supplementary features related to minor oscillations. This has been shown by exploring data, defining features and testing their discriminatory powers, on a large set of about 11000 GRF measurements related to stilettos, sport and patent leather shoes. The presented ideas can be used to develop classifiers able to differentiate between stilettos and other footwear, soft- and hard-sole shoes, shoes with and without heels.
In this paper, we propose a multichannel extension of Language-Queried Audio Source Separation (LASS) using Independent Vector Analysis (IVA). LASS enables the separation of arbitrary sound sources from a mixture based on natural language descriptions; however, conventional models assume a single-channel input and do not utilize spatial information. As one approach to extending LASS to multichannel processing, we consider utilizing LASS-separated signals as source models in Auxiliary-function-based IVA (AuxIVA), leveraging spatial and language query information for multichannel source separation. Furthermore, we investigate two approaches for integrating the LASS-separated signals into the IVA framework: a weighted arithmetic mean and a weighted geometric mean for mixing the source model variances. We demonstrate that the weighted geometric mean achieves higher separation performance through simulation experiments than the arithmetic mean. Experimental results indicate that the proposed method successfully extends LASS to multichannel source separation.
Maximum Likelihood (ML) Direction-of-Arrival (DoA) estimation under the Vectorized Covariance Matrix Model (VCMM) provides improved performance. However, the associated optimization problem remains computationally intractable due to its highly non-convex and multi-dimensional nature. To alleviate this issue, a sparse estimation strategy has been proposed, shown to be equivalent to the ML after a pre-whitening noise transformation when the regularization parameter is properly chosen. Yet, the resulting cost function remains challenging, with optimization limited to first-order methods such as the Proximal Gradient Algorithm (PGA), which suffers from slow convergence due to strong correlations in the dictionary matrix. This work exploits the decorrelation induced by the prewhitening transform to enable acceleration via a variable stepsize strategy. The transform increases the allowable stepsize by reducing the correlation of the dictionary near source directions, thus significantly improving convergence speed. Furthermore, the final iterations are shown to be equivalent with a more computationally demanding second-order algorithm, yielding an efficient approximation with reduced complexity. Numerical simulations confirm the predicted speed improvements in the case of 2D DoA estimation.
This paper introduces a novel service-oriented framework, Radio Propagation as a Service (RPaaS), that bridges the gap between raw sensor data and high-fidelity wireless channel simulations. RPaaS transforms noisy, sensor-derived point clouds into accurate 3D models through robust registration, segmentation, and edge detection. These models then feed into a GPU-accelerated ray tracing engine that computes multi-path propagation effects, while a separate module derives key electromagnetic and channel parameters. All components are orchestrated via a REST API in a Dockerized environment, enabling dynamic reconfiguration based on sensor data conditions. Experimental validation against commercial ray tracing tools and channel measurements demonstrates that our approach provides accurate simulations even in the presence of sensor noise.
The convex feasibility problem consists of finding a point in the intersection of closed convex sets. We propose a new type of algorithm to solve it in which randomly selected blocks of subgradient projectors are activated in parallel at each iteration and averaged through an extrapolation process. A key novelty is that the update is obtained by a random super relaxation step which allows for unbounded relaxation parameters, in sharp contrast with existing methods where they are deterministic and bounded by 2. Almost sure convergence is proved without requiring any regularity assumptions. Additionally, we establish meansquare convergence rates under linear regularity conditions. Numerical applications to signal and image recovery are demonstrated, which illustrate the benefits of super relaxations and random set activation.
Variability in histological staining protocols across medical centers introduces significant color differences. As a consequence, the performance of Computer-Aided Diagnosis (CAD) systems degrades when tested in images from unseen centers. To mitigate this issue, Blind Color Deconvolution (BCD) enables the separation of stain components. However, traditional analytical BCD methods are computationally expensive, while deep learning (DL)-based approaches offer efficient inference once trained. Despite their efficiency, existing DL-based methods lack effective priors to guide the learning process. In this work, we propose GSMNet, a DL-based model that incorporates a Gaussian Scale Mixture (GSM) prior to promote sparsity in concentration maps. To address the intractability introduced by this prior, we formulate an augmented probabilistic model that enables efficient variational inference. We evaluate GSMNet on two tasks - stain separation and breast cancer classification - demonstrating that it achieves performance superior to or comparable with state-of-the-art BCD methods while being significantly faster.
Are complex Deep Learning (DL) models necessary for effective heart sound segmentation? This paper demonstrates that an effective denoising approach combined with a standard lowcost probabilistic Markov-based segmentation method provides performance that is comparable to DL models while maintaining reasonable numerical complexity. The proposed pipeline takes advantage of the pseudo-periodicity of heart events across cardiac cycles for an efficient denoising step. This preprocessing step is reformulated as a constrained low-rank matrix inference in the framework of a graph signal processing. Denoised heart sounds are then fed to the classical Logistic Regression Hidden Semi-Markov Model (LR-HSMM) for segmentation purposes. A comparative study of the proposed pipeline with one of the most effective DL models, namely CNN-LSTM, using the CirCor dataset containing both healthy and murmur PCG signals was conducted to emphasize our purpose. Obtained results in terms of both segmentation of the two primary heart sounds S1 and S2 and murmur classification of this two-step pipeline are quasiequivalent to those obtained using the CNN-LSTM model, but with a relatively lower numerical complexity.
Epilepsy is a neurological disorder marked by recurrent and unpredictable disturbances in brain activity, leading to epileptic seizures (ES). One approach of diagnosing these seizures involves analyzing ElectroEncephaloGraphic (EEG) signals, which are highly dynamic and complex. Recently, researchers have been implementing Machine Learning (ML) Deep Learning (DL) techniques to perform automatic seizure detection of these EEG signals. ML models require extensive feature extraction techniques compared with DL ones. Despite the promising results of DL models in the detection task, their computational cost is still a challenge. Moreover, DL models require a large amount of training data which is not always available. Hence, in this paper, a hybrid channel-wise lightweight automatic seizure detection model based on vanilla autoencoders and K-NN classifier is introduced. The model is able to generate almost real-like data that can be used as data augmentation technique and increasing the dataset spread for DL models. Mean Square Error (MSE), Pearson Correlation Coefficient (PCC), and Power Spectral Density (PSD) similarity were used to evaluate the quality of the generated signal. Compared to other methods, this model provided a competitive average reconstruction loss of 0.4515, an average precision of 93.69%, and a low parameter count of 41.152k. The lightweight nature of this model could potentially be used in wearable clinical devices for real-time seizure detection.
Infant cry analysis serves as a non-invasive diagnostic tool for detecting pathological conditions. This study analyzes infant cry acoustics using formant distribution analysis in 3-D across Mel, Bark, Cochlear and Gammatone scales, visualizing fundamental frequency ($F_{0}$), and the first two formants ($F_{1}$), and ($F_{2}$) for normal vs. pathological cry classification. The dataset comprises cries from Baby Chillanto, Baby Chillanto 2.0, and DAIICT corpora, covering six pathologies. Harmonics-to-Noise Ratio (HNR), jitter, and shimmer are examined to assess phonatory irregularities, while violin plots highlight statistical dispersion in cry variability. Formant Space Area (FSA) via convex hull analysis quantifies spectral dispersion, offering a novel biomarker for pathology differentiation. For classification, both handcrafted spectral features (Mel Frequency Cepstral Coefficients, Gammatone Frequency Cepstral Coefficient, Cochlear Filter Cepstral Coefficients and Bark Frequency Cepstral Coefficients) and deep learning-based embeddings (HuBERT, wav2vec 2.0, XLS-R) are evaluated using CNN and Bi-LSTM models. BFCC outperforms MFCC, CFCC and GFCC, while deep learning embeddings enhance classification accuracy. Beyond classification, radar plots, convex hull mapping, and 3-D formant visualizations provide deeper insights into cry-based pathology detection.
Simulators are essential in autonomous systems research, providing a controlled test environment for self-driving vehicles, autonomous robots, and unmanned aerial vehicles. Despite recent significant improvements in simulation realism, the noticeable gap between the simulation and the real-world complexities persists, hindering the direct applicability of algorithms trained on simulated data to real-world scenarios. In this work, we extend the applicability of CARLA2Real 11https://github.com/stefanos50/CARLA2Real-a previously developed publicly available tool for enhancing CARLA's visual realism-to the latest Unreal Engine 5 version of the simulator, which already features improved rendering capabilities. Our approach utilizes a state-of-the-art image-to-image translation method to generate photorealism-enhanced autonomous driving-related visual data that target the characteristics of the real-world datasets, Cityscapes and KITTI. Based on this, we generated synthetic datasets from both the simulator and the photorealism enhancement model outputs, including their corresponding ground truth annotations for semantic segmentation. Subsequently, by employing the photorealism-improved synthetic data as training data, we conducted experiments to assess how the proposed approach affected the accuracy of a semantic segmentation approach. The findings illustrated that, although Unreal Engine 5 improves the baseline realism of CARLA, a sim2real gap still persists in various aspects of the scenes. However, our method significantly reduces this gap, leading to improved segmentation performance on real-world data.
This paper studies the adaptive detection of rangespread targets in a Gaussian environment, assuming a Kronecker-product structure for the disturbance covariance matrix. Invoking the principle of invariance, we identify a transformation group that significantly reduces the dimensionality of nuisance parameters, ensuring the constant false alarm rate (CFAR) property for all invariant statistics. A maximal invariant (MI) is derived, providing the foundation for new CFAR detectors that are invariant tests functionally depending on the MI. At the stage of detector design, two adaptive detectors are devised: the former employs the two-step strategy and incorporates the Kronecker maximum likelihood estimate based on secondary data, while the latter is the one-step generalized likelihood ratio test realized via an alternating-optimization algorithm. Both are invariant tests, and thus their CFAR properties with respect to the Kronecker covariance matrix are naturally guaranteed. Numerical results demonstrate the superior detection performance and robust CFAR behavior of both the detectors compared to conventional methods designed for the unstructured case.
Artificial Bandwidth Extension (ABE) enhances narrowband speech quality by reconstructing the lost high-frequency components essential for clarity and naturalness. In this work, we propose a novel ABE framework that integrates the constantQ Transform (CQT) and its variant within a lightweight neural network. Unlike traditional methods relying on the short-time Fourier transform (STFT), our approach leverages CQT's logarithmic frequency scaling and superior low-frequency resolution to better align with human auditory perception. Two CQTbased feature extraction schemes are introduced: a standard method that extracts narrowband (NB) CQT representations and a modified variant that employs a stacking and masking operation to compensate for missing high-frequency content. A compact Multi-Layer Perceptron (MLP) is then trained to map the extracted features to full wideband (WB) spectral representations. Phase reconstruction is achieved using either spectral folding or spectral shifting in conjunction with inverse CQT (iCQT), enabling effective reconstruction of the time-domain speech signal. Evaluations on the TIMIT dataset show that our model with modified CQT and spectral folding outperforms traditional methods, achieving lower Log Spectral Distance (LSD) and Visual Geometry Group (VGG) distance and higher Virtual Speech Quality Objective Listener (ViSQOL) values. Additionally, subjective evaluations using the MUSHRA framework validate the improvements in perceptual quality offered by the proposed approach.
This work focuses on joint image denoising and contour detection. On the one hand, contour detection on clean images (without noise) has been extensively studied from both variational formulation and deep learning perspectives. On the other hand, although the task of jointly denoising and contour detection has been largely considered in image processing literature using variational formulations, it has not been handled from the model-based neural network perspective. In this work, we propose an unfolded discrete Mumford-Shah procedure that enables us to bridge the gap between standard variational procedures designed to perform the combined denoising/edge detection task and black-box neural network designed for edge detection purpose.
In this paper, we describe an alternative approach to design the internal model control (IMC) principle for feedback active noise control (FBANC) by means of Laguerre filters. For this, we discuss the adequate choice of the Laguerre parameter and the filter order. It turns out that the Laguerre filter coefficients can be optimized in the same fashion as IMC designs that utilize finite impulse response (FIR) filters. Furthermore, when compared to FIR-based IMC designs, the proposed approach also has the advantage that significantly fewer filter coefficients need to be optimized. At the same time, the noise reduction performance is on par with higher order FIR-based designs. In terms of robust stability, our approach achieves good results in both nominal and non-nominal scenarios.
In conventional digital acquisition, a trade-off exists between allocating bits to capture a high dynamic range (HDR) signal and achieving high digital resolution (HDres). The Unlimited Sensing Framework (USF) overcomes this tradeoff by introducing modulo non-linearity before sampling and encoding HDR signals as low dynamic range folded samples. These samples are then decoded or unfolded using smoothness or parametric signal priors. The question investigated in this paper is whether we can simplify the reconstruction if we assume that the signals are sparse in some transform domain or dictionary. The key insight there is that even though the modulo non-linearity disrupts the sparsity in the original dictionary the output signal exhibits sparsity in a new dictionary derived from the original. Capitalizing on this, we propose a strategy that leverages sparse recovery techniques, such as Orthogonal Matching Pursuit (OMP) and Basis Pursuit (BP), to directly reconstruct sparse signals from modulo samples. In addition to numerical simulations, hardware experiments with a modulo ADC demonstrate the stability and robustness of our approach in real-world scenarios. Our work thus opens up a new direction of exploration, leveraging methods from the established field of sparse approximation for the emerging topic of USF.
