Department Publications

Test-time prompt tuning (TPT) has emerged as a promising technique for adapting large vision-language models (VLMs) to unseen tasks without relying on labeled data. However, the lack of dispersion between textual features can hurt calibration performance, which raises concerns about VLMs' reliability, trustworthiness, and safety. Current TPT approaches primarily focus on improving prompt calibration by either maximizing average textual feature dispersion or enforcing orthogonality constraints to encourage angular separation. However, these methods may not always have optimal angular separation between class-wise textual features, which implies overlooking the critical role of angular diversity. To address this, we propose A-TPT, a novel TPT framework that introduces angular diversity to encourage uniformity in the distribution of normalized textual features induced by corresponding learnable prompts. This uniformity is achieved by maximizing the minimum pairwise angular distance between features on the unit hypersphere. We show that our approach consistently surpasses state-of-the-art TPT methods in reducing the aggregate average calibration error while maintaining comparable accuracy through extensive experiments with various backbones on different datasets. Notably, our approach exhibits superior zero-shot calibration performance on natural distribution shifts and generalizes well to medical datasets. We provide extensive analyses, including theoretical aspects, to establish the grounding of A-TPT. These results highlight the potency of promoting angular diversity to achieve well-dispersed textual features, significantly …

Unmanned aerial vehicles are increasingly deployed across diverse domains including security surveillance, precision agriculture, infrastructure inspection, and logistics. However, their rapid proliferation raises critical concerns regarding unauthorized intrusions, smuggling, and adversarial exploitation. Robust drone detection and classification systems are therefore essential. In this paper, we propose a multi-modal framework that fuses frequency-modulated continuous-wave radar and acoustic sensing using a transformer encoder-based architecture. Unlike single-modality approaches, our method exploits cross-modal complementarity to improve robustness against environmental noise, false alarms, and stealthy drone operations. Notably, our approach directly processes raw acoustic waveforms without domain transformations such as spectrograms, reducing preprocessing overhead and improving real-time …

Intelligent control of RF transceivers adapting to dynamic operational conditions is essential in the modern and future communication systems. We propose a multi-agent neurosymbolic AI system, where AI agents are assigned for circuit components. Agents have an internal model and a corresponding control algorithm as its constituents. Modeling of the IF amplifier shows promising results, where the same approach can be extended to all the components, thus creating a fully intelligent RF system.

Semi-supervised Domain Generalization (SSDG) addresses the challenge of generalizing to unseen target domains with limited labeled data. Existing SSDG methods highlight the importance of achieving high pseudo-labeling (PL) accuracy and preventing model overfitting as the main challenges in SSDG. In this light, we show that the SSDG literature’s excessive focus on PL accuracy, without consideration for maximum data utilization during training, limits potential performance improvements. We propose a novel approach to the SSDG problem by aligning the intermediate features of our model with the semantically rich and generalized feature space of a Vision Language Model (VLM) in a way that promotes domain-invariance. The above approach is enhanced with effective image-level augmentation and output-level regularization strategies to improve data utilization and minimize overfitting. Extensive …

Identification of less-articulated objects using single-channel images, such as thermal images, is important in many applications, such as surveillance. However, in this domain, existing methods show poor performance due to high similarity among objects of the same category in the absence of color information (overlooking shape information) and de-emphasized texture information. Furthermore, variability in viewpoint adds more complexity as the features vary from side to side. We address these issues by constructing viewpoint-conditioned feature vectors and area-specific feature comparisons in separate feature spaces. These interventions enable leveraging the advancements of existing RGB-pre-trained ViT feature extractors while effectively adapting them to address the challenges specific to the thermal domain. We test our system with RGBNT100 (IR) vehicle dataset and a thermal maritime dataset acquired …

Spatial transcriptomics is an emerging field that enables the identification of functional regions based on the spatial distribution of gene expression. Integrating this functional information present in transcriptomic data with structural data from histopathology images is an active research area with applications in identifying tumor substructures associated with cancer drug resistance. Current histopathology-spatial-transcriptomic region segmentation methods suffer due to either making spatial transcriptomics prominent by using histopathology features just to assist processing spatial transcriptomics data or using vanilla contrastive learning that make histopathology images prominent due to only promoting common features losing functional information. In both extremes, the model gets either lost in the noise of spatial transcriptomics or overly smoothed, losing essential information. Thus, we propose our novel …

Semi-supervised domain generalization (SSDG) leverages a small fraction of labeled data alongside unlabeled data to enhance model generalization. Most of the existing SSDG methods rely on pseudo-labeling (PL) for unlabeled data, often assuming access to domain labels--a privilege not always available. However, domain shifts introduce domain noise, leading to inconsistent PLs that degrade model performance. Methods derived from semi-supervised learning baselines suffer particularly from lower PL accuracy, reducing the effectiveness of unlabeled data. We address this in a more challenging setting---domain-label agnostic SSDG---, where domain labels for unlabeled data are not available during training. First, we propose a feature modulation strategy that enhances class-discriminative features while suppressing domain-specific information. This modulation shifts features toward Similar Average Representations---a modified version of class prototypes---that are robust across domains, encouraging the classifier to distinguish between closely related classes and feature extractor to form tightly clustered, domain-invariant representations. Second, to mitigate domain noise and improve pseudo-label accuracy, we introduce a loss-scaling function that dynamically lowers the fixed confidence threshold for pseudo-labels, optimizing the use of unlabeled data. With these key innovations, our approach achieves significant improvements on four major domain generalization benchmarks--even without domain labels. We will make the code available.

Neonatal jaundice, caused by elevated bilirubin levels, requires timely diagnosis to prevent severe complications. Existing deep learning approaches primarily focus on binary classification of jaundice from visual data, but do not provide quantitative bilirubin estimation. A binary decision alone is insufficient for clinical management, where treatment thresholds depend on precise bilirubin levels. This paper presents a deep-learning-based contactless framework for neonatal jaundice assessment using image-based analysis. The proposed lightweight multi-stage pipeline integrates eye-region obfuscation, skin segmentation, and deep-learning-based image classification and bilirubin estimation. The system first performs jaundice classification; upon positive detection, a regression model is invoked to estimate bilirubin. Furthermore, image-based neonatal screening raises privacy concerns, as facial images of …

In coastal environments, detecting and classifying low radar cross-section (RCS) moving targets such as dinghies and trawlers using pulse radar system is challenging due to strong sea and land clutter. The clutter reduces detection accuracy. Further, the detection and classification heavily depend on radar operator expertise. To overcome these limitations, we propose an operator-independent framework for detection and classification. This paper presents a hybrid approach that combines Robust Principal Component Analysis (RPCA) with deep learning-based object detection models to improve target detection using radar Plan Position Indicator (PPI) images. RPCA is first applied as a low rank nature clutter separation from sparse moving targets. Detection is then performed using You Only Look Once (YOLO) and Detection Transformer (DETR) models on these clutter-suppressed representations. Experimental …

Civilian GNSS receivers are vulnerable to spoofing due to open signal structures, while adaptive electronic-warfare platforms increasingly synthesize high-fidelity counterfeit signals. Existing detection methods depend on handcrafted features or supervised learning, limiting robustness under distribution shifts and novel attacks. This paper introduces NEMESIS, a self-supervised framework based on joint embedding predictive architectures (JEPA) in the wavelet time-frequency domain, which leverages unlabeled RF corpora via masked time-frequency prediction to learn spoofing-relevant embeddings directly from raw RF signals without labels or contrastive augmentation. Experiments demonstrate that wavelet-domain JEPA substantially outperforms time- and frequency-domain variants and achieves competitive accuracy relative to contrastive learning while significantly reducing labeling requirements.

Fire is a serious threat to both human life and property, and fire safety is crucial. Many traditional methods have been developed to detect fires. However, inspired by past research that used Wi-Fi signals for monitoring environments and recognizing activities, this approach aims to use Wi-Fi signals to detect fire that utilize existing wireless communication systems for non-intrusive sensing. The idea is that fire and smoke can interfere with Wi-Fi signals passing through an environment, then Wi-Fi CSI could provide detailed information about the wireless channel between a Wi-Fi transmitter and a receiver. This research demonstrates the effectiveness of using Wi-Fi for fire detection, utilizing Raspberry Pi and ESP32 devices with ESP32 CSI tool. It analyses the temporal shift in Wi-Fi CSI Amplitude before, during, and after the ignition of a flame. The results showed that fire causes a noticeable variation in the CSI …

This paper presents a multi agent neurosymbolic artificial intelligence architecture to dynamically adapt operating parameters of radio frequency (RF) transceivers. The proposed dual branch architecture combines deep learning feature extraction with deterministic symbolic reasoning, leveraging sensed physical parameters such as signal power and spectrum occupancy, to ensure self-adapting and robust hardware safety. Component specific control agents are trained using digital twins and integrated with physical hardware via shared memory space and socket-based inter-process communication for low latency. Experimental results demonstrate accurate modeling and rapid parameter adaptation for error vector magnitude optimization, providing a scalable foundation for intelligent self-adapting RF transceivers.

Sensing and early detection of small unmanned aerial systems (sUAS) are critically important in modern-day defense. In dense urban and indoor environments, detection becomes extremely challenging due to dense multipath, fading, low-altitude flight, and non-line-of-sight (NLOS) radio-frequency propagation. This paper presents a continuous-wave multiple-input multiple-output radar and a deep learning model for sUAS detection using NLOS signals. The radar operates at 2.47 GHz, and spectral correlation densities derived from rotational micro-Doppler signatures from the rotor blades are used as inputs to the deep learning model. Experimental results demonstrate an overall detection accuracy of across a dataset of five drone types, confirming the feasibility of sUAS detection in dense urban environments without direct line-of-sight conditions.

This paper presents the design and implementation of a multi-band multi-mode radar for reliable detection and identification of micro unmanned aerial systems (UASs). A rotational micro-Doppler continuous-wave radar operating at 5.85 GHz and a frequency-modulated continuous-wave (FMCW) radar operating with 77–80 GHz bandwidth are used for multi-modal radio frequency (RF) sensing, localization and detection. Fusion of the two radars is achieved with a neurosymbolic artificial intelligent model. Cyclostationary feature extraction is used for the micro-Doppler signature extraction. Wavelet domain feature extraction is used on the FMCW radar outputs. A custom neural engine is used for multi-modal sensor fusion across the two RF bands and radar types.

Low-resolution analog-to-digital converters (ADCs) are a promising way to cut power consumption in future 6G receivers. However, they break many assumptions behind traditional communication system design. Instead of optimizing each block separately, this work uses an end-to-end autoencoder (AE) to jointly learn and optimize the transmitter and receiver for systems with phase quantization. Two key obstacles arise: (1) phase quantization is non-differentiable, which blocks gradientbased training back to the transmitter, and (2) standard crossentropy loss leads to poor block error rate (BLER) performance. To overcome these issues, we introduce a differentiable approximation of phase quantization and a new loss function that adds a distance-based regularization term to improve symbol separability. Simulations show that the proposed AE achieves near-optimal BLER for uncoded -ary phase shift keying …

This paper proposes a support vector machine (SVM)-based framework for signal detection in low-resolution single-input multiple-output (SIMO) systems, operating without channel state information (CSI). We consider n-bit phase quantization under M-PSK modulation and develop a CSI-free, model-free detector that directly maps quantized observations to symbol estimates. For QPSK modulation, the proposed method achieves symbol error probability (SEP) performance within 15% of the coherent maximum likelihood detector (MLD). We further extend the framework to higher-order M-PSK modulations, providing the first model-free, CSI-free SEP performance characterization for low-resolution, phase-quantized SIMO systems; addressing a regime where no general MLD solution is available.

Integrated sensing and communication (ISAC) is expected to be a key enabler for future wireless networks, improving spectral and hardware efficiency by jointly performing radar sensing and wireless communication within a unified framework. This paper proposes BeamNet , an unsupervised deep learning framework for transmit beamforming in dual-function radar-communication systems operating over general fading with imperfect channel state information (CSI). BeamNet maps noisy estimates of the communication and sensing channels to a transmit beamforming vector and is trained end-to-end by maximizing a weighted sum of the communication rate (CR) and sensing rate (SR), thereby learning the CR–SR Pareto frontier without beamforming labels or embedded optimization solvers. Using Rayleigh fading with perfect CSI, we first show that BeamNet reproduces the analytical Pareto-optimal beamforming solutions. We then use BeamNet to characterize, for Nakagami- m and Rician fading, the CR–SR trade-off across a range of fading parameters, and to assess robustness under distribution mismatch between training and test channels. Finally, under imperfect CSI, we demonstrate that BeamNet yields CR–SR trade-offs that are consistently sandwiched between the perfect-CSI and mismatched analytical baselines, outperforming the closed-form beamformer applied to imperfect CSI and recovering part of the performance loss caused by channel estimation errors. These results indicate that unsupervised learning offers a flexible and robust approach to ISAC beamforming in fading environments with imperfect channel knowledge.

Diabetes is associated with progressive microvascular remodelling, commonly assessed using retinal imaging, yet alternative non-invasive vascular biomarkers remain underexplored. Conjunctival vessels are directly visible and may reflect systemic microvascular changes; however, reliable quantification is challenged by the need to preserve vascular topology during segmentation. In this work, we introduce a curated conjunctival imaging dataset with diameter-specific vessel annotations and demonstrate that topology-preserving vessel extraction is essential for robust tortuosity estimation. We develop a deep learning framework that prioritises vascular continuity and geometric fidelity while selectively segmenting clinically relevant vessel calibres, integrating multi-scale contextual modelling and attention-guided feature fusion to reduce fragmentation without post-processing. Compared with multiple state-of-the-art …

The Relative Transfer Matrix (ReTM), recently introduced as a generalization of the relative transfer function for multiple receivers and sources, shows promising performance when applied to speech enhancement in noisy environments. Estimating the ReTM of sound sources by exploiting the covariance matrices of multichannel recordings is highly beneficial for practical applications and, to date, remains the only proposed approach. This paper investigates deep learning-based ReTM estimation. We propose three novel supervised learning frameworks using time and short-time frequency transform domain convolutional networks, and a Long Short-Term Memory-based recurrent neural network. Experimental results demonstrate that the proposed models achieve more accurate estimation of the ReTM using five objective metrics compared to the covariance-based method. We also show the effectiveness of the proposed frameworks for speech enhancement, achieving performance on par with the baseline method.

Isolating a desired speech signal in noisy multi-talker conversational scenarios is a key requirement for augmented reality (AR) wearable microphone array systems. In this work, a binaural target speaker extraction (TSE) framework, termed BiTSE, is proposed. It leverages both spatial and temporal cues, specifically the direction-of-arrival (DoA) of the target speaker and corresponding voice activity information, to guide the extraction process. Built upon a binaural signal denoising architecture, our model integrates three key enhancements: (i) a DoA-aware attention mechanism using cyclic positional embeddings, (ii) a timestamp-based masking strategy that utilizes speaker activity to suppress non-target segments, and (iii) a novel two-stage loss optimization strategy that first trains the model for robust denoising and then fine-tunes it to improve perceptual quality. Evaluations on the SPeech Enhancement for Augmented Reality (SPEAR) challenge dataset demonstrate that the proposed BiTSE consistently improves upon conventional approaches, leading to enhanced signal fidelity and perceptual quality.

Test-time prompt tuning (TPT) has emerged as a promising technique for adapting large vision-language models (VLMs) to unseen tasks without relying on labeled data. However, the lack of dispersion between textual features can hurt calibration performance, which raises concerns about VLMs' reliability, trustworthiness, and safety. Current TPT approaches primarily focus on improving prompt calibration by either maximizing average textual feature dispersion or enforcing orthogonality constraints to encourage angular separation. However, these methods may not always have optimal angular separation between class-wise textual features, which implies overlooking the critical role of angular diversity. To address this, we propose A-TPT, a novel TPT framework that introduces angular diversity to encourage uniformity in the distribution of normalized textual features induced by corresponding learnable prompts …

Material that constitute aggregate depend on their packing efficiency that is defined by the shape distribution of aggregates. Geometrical shape descriptors would not capture morphological aspects of different dimensional scales and frequencies, failing to wholistically numerically represent shape in material prediction models. Fourier shape descriptors are used to analyse single particle morphology, but efficiency to characterize shape of lump samples has not been assessed. This study analyses aggregates milled for different number of revolutions (0, 200, 500 and 1000) in a Los Angeles Abrasion Value instrument, that contrives morphological alterations of different scales. Means of zonal frequency components amplitudes (form, angularity and texture) are statistically different for all classes with more than 99 % confidence. Support Vector Machine and K Nearest Neighbour algorithms classified pairwise …

The emergence of 5G technology has transformed network architecture, with network slicing technology standing out as a crucial feature for managing diverse and demanding applications. This innovative technology allows the creation of multiple virtual networks within a single physical infrastructure, each dedicated to specific needs such as latency, bandwidth, and reliability. These capabilities are vital for supporting services like enhanced mobile broadband, ultra-reliable low-latency communications, and massive machine-type communications. Predicting the optimal configuration of network slices is challenging due to the complex and dynamic nature of 5G environments. While traditional methods often lack the necessary accuracy, machine learning models have shown promise in addressing this challenge. However, the adoption of these models introduces the challenge of interpretability, as understanding how …

This manuscript introduces NeuroDetect, a model-free deep learning-based signal detection framework tailored for phase-modulated wireless systems with low-resolution analog-to-digital converters (ADCs). The proposed framework eliminates the need for explicit channel state information, which is typically difficult to acquire under coarse quantization. NeuroDetect utilizes a neural network architecture to learn the nonlinear relationship between quantized received signals and transmitted symbols directly from data. It achieves near-optimum performance, within a worst-case 12% margin of the maximum likelihood detector that assumes perfect channel knowledge. We rigorously investigate the interplay between ADC resolution and detection accuracy, introducing novel penalty metrics that quantify the effects of both quantization and learning errors. Our results shed light on the design trade-offs between ADC resolution and detection accuracy, providing future directions for developing energy-efficient high-speed and wideband wireless systems.

Image-to-point cross-modal learning has emerged to address the scarcity of large-scale 3D datasets in 3D representation learning. However, current methods that leverage 2D data often result in large, slow-to-train models, making them computationally expensive and difficult to deploy in resource-constrained environments. The architecture design of such models is therefore critical, determining their performance, memory footprint, and compute efficiency. The Joint-embedding Predictive Architecture (JEPA) has gained wide popularity in self-supervised learning for its simplicity and efficiency, but has been under-explored in cross-modal settings, partly due to the misconception that masking is intrinsic to JEPA. In this light, we propose CrossJEPA, a simple Cross-modal Joint Embedding Predictive Architecture that harnesses the knowledge of an image foundation model and trains a predictor to infer embeddings of specific rendered 2D views from corresponding 3D point clouds, thereby introducing a JEPA-style pretraining strategy beyond masking. By conditioning the predictor on cross-domain projection information, CrossJEPA purifies the supervision signal from semantics exclusive to the target domain. We further exploit the frozen teacher design with a one-time target embedding caching mechanism, yielding amortized efficiency. CrossJEPA achieves a new state-of-the-art in linear probing on the synthetic ModelNet40 (94.2%) and the real-world ScanObjectNN (88.3%) benchmarks, using only 14.1M pretraining parameters (8.5M in the point encoder), and about 6 pretraining hours on a standard single GPU. These results position CrossJEPA …