Low reliability has consistently been a challenge in the application of deep learning models for high risk decision-making scenarios. In medical image segmentation, for instance, multiple expert annotations can be consulted to reduce subjective bias and reach a consensus, thereby enhancing the segmentation accuracy and reliability. To develop a reliable lesion segmentation model, we leverage the uncertainty introduced by multiple annotations, enabling the model to better capture real-world diagnostic variability and provide more informative predictions. Since a reliable model should produce calibrated uncertainty estimates that align with actual predictive performance, we propose CalDiff, a novel framework designed to calibrate model uncertainty in lesion segmentation and mitigate the risk of overconfident yet incorrect predictions. To harness the superior generative ability of diffusion mod els, a dual step-wise and sequence-aware calibration mechanism is proposed on the basis of the sequential nature of diffusion models. We evaluate the calibrated model through a comprehensive quantitative and visual analysis, thus ad dressing the previously overlooked challenge of assessing uncertainty calibration and model reliability in scenarios with multiple annotations and multiple predictions. Experimental results on two multi-annotated lesion segmentation datasets demonstrate that CalDiff produces uncertainty maps that can reflect informative low confidence areas, which can further indicate the false predictions potentially made by the model. By calibrating the uncertainty in the training phase, the uncertain areas produced from our model are more closely correlated with areas where the model has made errors in the inference. In summary, the uncertainty captured by our CalDiff framework can serve as apowerful indicator, which can help mitigate the risks of adopting model's outputs, allowing clinicians to prioritize reviewing areas or slices with higher uncertainty and enhancing the model's reliability and trustworthiness in real clinical practice.

As large language models continue to grow in size, parameter-efficient fine-tuning (PEFT) has become increasingly crucial.While lowrank adaptation (LoRA) offers a solution through low-rank updates, its static rank allocation may yield suboptimal results.Adaptive low-rank adaptation (AdaLoRA) improves this with dynamic allocation but remains sensitive to initial and target rank configurations.We introduce AROMA, a framework that automatically constructs layer-specific updates by iteratively building up rank-one components with very few trainable parameters that gradually diminish to zero.Unlike existing methods that employ rank reduction mechanisms, AROMA introduces a dual-loop architecture for rank growth.The inner loop extracts information from each rank-one subspace, while the outer loop determines the number of rankone subspaces, i.e., the optimal rank.We reset optimizer states to maintain subspace independence.AROMA significantly reduces parameters compared to LoRA and AdaLoRA while achieving superior performance on natural language understanding and generation, commonsense reasoning, offering new insights into adaptive PEFT.

Abstract. The classical Mumford–Shah (MS) model has been successful in some medical image segmentation tasks, providing segmentation results with smooth boundaries of objects. However, the MS model, which operates at the pixel level of the images, faces challenges when dealing with medical images with low contrast or unclear edges. In this paper, we begin by using a feature extractor to capture high-dimensional deep features that contain more comprehensive semantic information than pixel-level data alone. Inspired by the MS model, we develop a variational model that incorporates threshold dynamics (TD) regularization for segmenting each feature. We obtain the final segmentation result for the original image by assembling segmentation results of all the features. This process results in MS-MGNet, a lightweight trainable segmentation network with a similar architecture to many encoder–decoder networks. The intermediate layers of MS-MGNet are designed by unrolling the numerical scheme based on the multigrid method for solving the variational model. We provide interpretability for the encoder–decoder architecture by elucidating the roles of each layer and offering explanations of the underlying mathematical models. By incorporating the TD regularizer, we integrate spatial priors from the variational models into the network architecture, resulting in better segmentation results with smoother edges and a certain robustness to noise. Compared to some relevant methods, experimental results on the selected data sets with low contrast or unclear edges show that the proposed method can achieve better segmentation performance with fewer parameters, even when trained on smaller data sets.

Syntax-Guided Synthesis (SyGuS) is a general solving framework for program synthesis. Previous researchers have proposed a divide-and-conquer strategy to divide the grammar search space in order to search separately. This framework converts the program synthesis problem into a decision tree classification problem. It unifies the obtained partial results into a complete program by learning decision trees according to conditional expression statements. However, due to the unknowns and uncertainties of sample label categories and sample attribute features in program synthesis, as well as the limitation of traditional decision tree learning methods to immediate gains, it results in long solving times and the large size of the candidate program. To address the above problems, we propose a deep reinforcement learning-based synthesis strategy for syntax-guided programs: a) the decision tree learning process is modeled as a Markov chain decision process, addressing the learning of knowledge patterns under conceptual drift and the weighting of the importance of sample attribute features; b) a graph neural network is used to extract the feature information of the sample attributes as action embedding feature values; (c) a reward function is proposed to evaluate the classification accuracy of the sample attributes as a feedback mechanism. We have implemented our approach in a tool called RLSolver. According to the experiments, the solving times ofRLSolver are comparable to EUSolver on small-scale task datasets. RLSolver substantially reduces the solving times on large-scale task datasets, solving two more SyGus tasks than EUSolver in the same solving time limit. In addition, RLSolver reduces the number of decision tree learning rounds by 60% and the size of decision tree size by 40% after using the multi-round model training strategy and early-stop strategy.

Activity recognition is an important component of many pervasive computing applications. Radio-based activity recognition has the advantage that it does not have the privacy concern compared with camera-based solutions, and subjects do not have to carry a device on them. It has been shown channel state information (CSI) can be used for activity recognition in a device-free setting. With the proliferation of wireless devices, it is important to understand how radio frequency interference (RFI) can impact on pervasive computing applications. In this article, we investigate the impact of RFI on device-free CSI-based location-oriented activity recognition. We present data to show that RFI can have a significant impact on the CSI vectors. In the absence of RFI, different activities give rise to different CSI vectors that can be differentiated visually. However, in the presence of RFI, the CSI vectors become much noisier, and activity recognition also becomes harder. Our extensive experiments show that the performance may degrade significantly with RFI. We then propose a number of countermeasures to mitigate the impact of RFI and improve the performance. We are also the first to use complex-valued CSI along with the state-of-the-art Sparse Representation Classification method to enhance the performance in the environment with RFI.

Cross-correlation is a popular signal processing technique used in numerous location tracking systems for obtaining reliablerangeinformation. However, its efficient design and practical implementation has not yet been achieved on mote platforms that are typical in wireless sensor network due to resource constrains. In this paper, we proposeStructS-XCorr: cross-correlation via structured sparse representation, a new computing framework for ranging based on$\ell _1$-norm minimization[1]and structured sparsity. The key idea is to compress the ranging signal samples on the mote by efficient random projections and transfer them to a central device; where a convex optimization process estimates the range by exploiting the sparse signal structure in the proposedcorrelationdictionary. Through theoretical validation, extensive empirical studies and experiments on anend-to-endacoustic ranging system implemented on resource limited off-the-shelf sensor nodes, we show that the proposed framework can achieve up totwo orders of magnitudebetter performance compared to other approaches such as working on DCT domain and downsampling. Compared to the standard cross-correlation, it is able to obtain range estimates with a bias of 2-6 cm with 30 percent and approximately 100 cm with 5 percent compressed measurements. Its structured sparsity model is able to improve the ranging accuracy by 40 percent under challenging recovery conditions (such as high compression factor and low signal-to-noise ratio) by overcoming limitations due to dictionary coherence.

Face recognition is one of the important components of many smart devices apps, e.g., face unlocking, people tagging and games on smart phones, tablets, or smart glasses. Sparse Representation Classification (SRC) is a state-of-the-art face recognition algorithm, which has been shown to outperform many classical face recognition algorithms in OpenCV, e.g., Eigenface algorithm. The success of SRC is due to its use of 21 optimization, which makes SRC robust to noise and occlusions. Since 21 optimization is computationally intensive, SRC uses random projection matrices to reduce the dimension of the 21 problem. However, random projection matrices do not give consistent classification accuracy as they ignored the prior knowledge of the training set. In this paper, we propose to exploit the prior knowlege of the training set to improve the recognition accuracy. It first learns the optimized projection matrix from the training set to produce consistent recognition performance then applies 21-based classification based on the group sparsity structure of SRC to further improve the recognition accuracy. Our evaluations, based on publicly available databases and real experiment, show that face recognition using optimized projection matrix is 8-17 percent more accurate than its random counterpart and Eigenface algorithm, and the recognition accuracy can be further improved by up to 5 percent by exploiting group sparsity structure. Furthermore, the optimized projection matrix does not have to be re-calculated even if new faces are added to the training set. We implement the SRC with optimized projection matrix on Android smartphones and find that the computation of residuals in SRC is a severe bottleneck, taking up 85-90 percent of the computation time. To address this problem, we propose a method to compute the residuals approximately, which is 50 times faster with little sacrificing recognition accuracy. Lastly, we demonstrate the feasibility of our new algorithm by the implementation and evaluation of a new face unlocking app and show its robustness to variation of poses, facial expressions, lighting changes, and occlusions.

Real-time target tracking is an important service provided by embedded camera networks. The first step in target tracking is to extract the moving targets from the video frames, which can be realised by using background subtraction. For a background subtraction method to be useful in embedded camera networks, it must be both accurate and computationally efficient because of the resource constraints on embedded platforms. This makes many traditional background subtraction algorithms unsuitable for embedded platforms because they use complex statistical models to handle subtle illumination changes. These models make them accurate but the computational requirement of these complex models is often too high for embedded platforms. In this paper, we propose a new background subtraction method which is both accurate and computationally efficient. We propose a baseline version which uses luminance only and then extend it to use colour information. The key idea is to use random projection matrics to reduce the dimensionality of the data while retaining most of the information. By using multiple datasets, we show that the accuracy of our proposed background subtraction method is comparable to that of the traditional background subtraction methods. Moreover, to show the computational efficiency of our methods is not platform specific, we implement it on various platforms. The real implementation shows that our proposed method is consistently better and is up to six times faster, and consume significantly less resources than the conventional approaches. Finally, we demonstrated the feasibility of the proposed method by the implementation and evaluation of an end-to-end real-time embedded camera network target tracking application.

Activity recognition is an important component of many pervasive computing applications. Device-free activity recognition has the advantage that it does not have the privacy concern of using cameras and the subjects do not have to carry a device on them. Recently, it has been shown that channel state information (CSI) can be used for activity recognition in a device-free setting. With the proliferation of wireless devices, it is important to understand how radio frequency interference (RFI) can impact on pervasive computing applications. In this paper, we investigate the impact of RFI on device-free CSI-based location-oriented activity recognition. We conduct experiments in environments without and with RFI. We present data to show that RFI can have a significant impact on the CSI vectors. In the absence of RFI, different activities give rise to different CSI vectors that can be differentiated visually. However, in the presence of RFI, the CSI vectors become much noisier and activity recognition also becomes harder. Our extensive experiments shows that the performance of state-of-the-art classification methods may degrade significantly with RFI. We then propose a number of counter measures to mitigate the impact of RFI and improve the location-oriented activity recognition performance. Our evaluation shows the proposed method can improve up to 10% true detection rate in the presence of RFI. We also study the impact of bandwidth on activity recognition performance. We show that with a channel bandwidth of 20 MHz (which is used by WiFi), it is possible to achieve a good activity recognition accuracy when RFI is present.

The information-preserving sampling properties of compressive sensing have found a number of successful applications, such as sensor scheduling, localisation and tracking to deal with the resource constraints of the embedded systems. In this paper, we investigate an approach to improve the performance of compressive sensing applications through a novel strategy for optimising the projection matrix. We formulate the projection matrix optimisation problem and apply greedy algorithm to solve the optimisation problem efficiently. We evaluate the proposed approach by an emerging background subtraction method designed specifically for the embedded systems and show the proposed approach outperforms existing approaches significantly with little overhead.

Acoustic Sensor Networks (ASNs) have a wide range of applications in natural and urban environment monitoring, as well as indoor activity monitoring. In-network classification is critically important in ASNs because wireless transmission costs several orders of magnitude more energy than computation. The main challenges of in-network classification in ASNs include effective feature selection, intensive computation requirement and high noise levels. To address these challenges, we propose a sparse representation based feature-less, low computational cost, and noise resilient framework for in-network classification in ASNs. The key component of Sparse Approximation based Classification (SAC), ℓ1 minimization, is a convex optimization problem, and is known to be computationally expensive. Furthermore, SAC algorithms assumes that the test samples are a linear combination of a few training samples in the training sets. For acoustic applications, this results in a very large training dictionary, making the computation infeasible to be performed on resource constrained ASN platforms. Therefore, we propose several techniques to reduce the size of the problem, so as to fit SAC for in-network classification in ASNs. Our extensive evaluation using two real-life datasets (consisting of calls from 14 frog species and 20 cricket species respectively) shows that the proposed SAC framework outperforms conventional approaches such as Support Vector Machines (SVMs) and k-Nearest Neighbor (kNN) in terms of classification accuracy and robustness. Moreover, our SAC approach can deal with multi-label classification which is common in ASNs. Finally, we explore the system design spaces and demonstrate the real-time feasibility of the proposed framework by the implementation and evaluation of an acoustic classification application on an embedded ASN testbed.

Cross-correlation is a popular signal processing technique used in numerous localization and tracking systems for obtaining reliable range information. However, a practical efficient implementation has not yet been achieved on resource constrained wireless sensor network platforms. We propose cross-correlation via sparse representation: a new framework for ranging based on l1-minimization. The key idea is to compress the signal samples on the mote platform by efficient random projections and transfer them to a central device, where a convex optimization process estimates the range by exploiting its sparsity in our proposed correlation domain. Through sparse representation theory validation, extensive empirical studies and experiments on an end-to-end acoustic ranging system implemented on resource limited off-the-shelf sensor nodes, we show that the proposed framework, together with the proposed correlation domain achieved up to two order of magnitude better performance compared to naive approaches such as working on DCT domain and downsampling. Furthermore, compared to cross-correlation results, 30-40% measurements are sufficient to obtain precise range estimates with an additional bias of only 2-6cm for high accuracy application requirements, while 5% measurements are adequate to achieve approximately 100cm precision for lower accuracy applications.

Background subtraction is often the first step in many computer vision applications such as object localisation and tracking. It aims to segment out moving parts of a scene that represent object of interests. In the field of computer vision, researchers have dedicated their efforts to improve the robustness and accuracy of such segmentations but most of their methods are computationally intensive, making them non-viable options for our targeted embedded camera platform whose energy and processing power is significantly more constrained. To address this problem as well as maintain an acceptable level of performance, we introduce Compressive Sensing (CS) to the widely used Mixture of Gaussian to create a new background subtraction method. The results show that our method not only can decrease the computation significantly (a factor of 7 in a DSP setting) but remains comparably accurate.

Sparse approximation has now become a buzzword for classification in numerous research domains. We propose a distributed sparse approximation method based on l1 minimization for frog sound classification, which is tailored to the resource constrained wireless sensor networks. Our pilot study demonstrates that l1 minimization can run on wireless sensor nodes producing satisfactory classification accuracy.

Cross-correlation is a popular signal processing technique used for obtaining reliable range information. Recently, a practical and efficient implementation of cross-correlation (via sparse approximation) was demonstrated on resource constrained wireless sensor network platforms, where the key idea was to compress the received signal samples, and transfer them to central device where the range information was retrieved by l1-minimization. Although, this mechanism yields accurate ranging results, its applicability is limited due to its slow execution speed and inaccurate recovery of the correlation peak magnitude, which implicitly provides the useful measure of signal-to-noise ratio. In this work, we propose Fast Gradient Projection (F-GP), a new l1-minimization algorithm, which overcomes the existing limitations, and provides fast and accurate ranging.

Background subtraction is often the first step of many computer vision applications. For a background subtraction method to be useful in embedded camera networks, it must be both accurate and computationally efficient because of the resource constraints on embedded platforms. This makes many traditional background subtraction algorithms unsuitable for embedded platforms because they use complex statistical models to handle subtle illumination changes. These models make them accurate but the computational requirement of these complex models is often too high for embedded platforms. In this paper, we propose a new background subtraction method which is both accurate and computational efficient. The key idea is to use compressive sensing to reduce the dimensionality of the data while retaining most of the information. By using multiple datasets, we show that the accuracy of our proposed background subtraction method is comparable to that of the traditional background subtraction methods. Moreover, real implementation on an embedded camera platform shows that our proposed method is at least 5 times faster, and consumes significantly less energy and memory resources than the conventional approaches. Finally, we demonstrated the feasibility of the proposed method by the implementation and evaluation of an end-to-end real-time embedded camera network target tracking application.