We propose a new framework for Weakly Supervised Object Detection (WSOD), a domain that traditionally relies on image-level labels. In addressing the inherent limitations of current WSOD methods, particularly their reliance on image-level annotations that result in inaccurate bounding box selections, we develop a framework that iteratively utilizes weak supervision and refines it to progressively enhance the supervision of the object detector throughout the training process. Specifically, we employ the Segment Anything Model (SAM) to generate initial pseudo-labels bounding boxes from the point prompts generated by Class Activation Mapping (CAM). Our approach tackles the challenge of label noise, where pseudo-labels bounding boxes might only capture parts of objects. We enhance our ability to distinguish between complete and partial detected objects by leveraging an instance-dependent, particularly part-based noise correction model. Our method is inspired by learning methods focusing on part-based representations for object detection and recognition, as well as from human perception, which typically simplifies complex visual information into simpler, constituent parts. Our experiments, conducted in various settings beyond WSOD, including Semi-Supervised Object Detection (SSOD) and Weakly Supervised Instance Segmentation (WSIS), validate the efficacy of our approach.

The problem of fully supervised classification is that it requires a tremendous amount of annotated data, however, in many datasets a large portion of data is unlabeled. To alleviate this problem semi-supervised learning (SSL) leverages the knowledge of the classifier on the labeled domain and extrapolates it to the unlabeled domain which has a supposedly similar distribution as annotated data. Recent success on SSL methods crucially hinges on thresholded pseudo labeling and thereby consistency regularization for the unlabeled domain. However, the existing methods do not incorporate the uncertainty of the pseudo labels or unlabeled samples in the training process which are due to the noisy labels or out of distribution samples owing to strong augmentations. Inspired by the recent developments in SSL, our goal in this paper is to propose a novel unsupervised uncertainty-aware objective that relies on aleatoric and epistemic uncertainty quantification. Complementing the recent techniques in SSL with the proposed uncertainty-aware loss function our approach outperforms or is on par with the state-of-the-art over standard SSL benchmarks while being computationally lightweight. Our results outperform the state-of-the-art results on complex datasets such as CIFAR-100 and Mini-ImageNet.

In this letter, we present DeDRLSSL: a generic semisupervised noise suppression framework. The proposed model is based on a reinforcement learning system for learning contrastive features to refine the features utilized in consistency matching for semisupervised learning (SSL). The proposed method outperforms the state-of-the-art supervised models in terms of error compensation for Inertial Measurement Unit data from various evaluation metrics and improves the baselines for yaw estimation on average by 38% and 28% across the benchmarks for 30% and 50% of labeled data, respectively. Our approach can be adapted to any SSL approach to compensate for the problem of label scarcity.

A Machine Learning approach has been developed to predict iron threshold exceedances in sub-regions of a drinking water distribution network from data collected the previous year. Models were trained using parameters informed by Self-Organising Map analysis based on ten years of water quality sampling data, pipe data and discolouration customer contacts from a UK network supplying over 2.3 million households. Twenty combinations of input parameters (network conditions) and three learning algorithms (Random Forests, Support Vector Machines and RUSBoost Trees) were tested. The best performing model was found to be Random Forests with input parameters of iron, turbidity, 3-day Heterotrophic Plate Counts, and high priority dead ends per District Metered Area. Different exceedance levels were tested and prediction accuracies of above 70% were achieved for UK regulatory concentration of 200 µg/L. Predicted probabilities per network sub-region were used to provide relative risk ranking to inform proactive management and investment decisions.

Proximal gradient method has a major role in solving nonsmooth composite optimization problems. However, in some machine learning problems related to black-box optimization models, the proximal gradient method could not be leveraged as the derivation of explicit gradients are difficult or entirely infeasible. Several variants of zeroth-order (ZO) stochastic variance reduced such as ZO-SVRG and ZO-SPIDER algorithms have recently been studied for nonconvex optimization problems. However, almost all the existing ZO-type algorithms suffer from a slowdown and increase in function query complexities up to a small-degree polynomial of the problem size. In order to fill this void, we propose a new analysis for the stochastic gradient algorithm for optimizing nonconvex, nonsmooth finite-sum problems, called ZO-PSVRG+ and ZO-PSPIDER+. The main goal of this work is to present an analysis that brings the convergence analysis for ZO-PSVRG+ and ZO-PSPIDER+ into uniformity, recovering several existing convergence results for arbitrary minibatch sizes while improving the complexity of their ZO oracle and proximal oracle calls. We prove that the studied ZO algorithms under Polyak-Łojasiewicz condition in contrast to the existent ZO-type methods obtain a global linear convergence for a wide range of minibatch sizes when the iterate enters into a local PL region without restart and algorithmic modification. The current analysis in the literature is mainly limited to large minibatch sizes, rendering the existing methods unpractical for real-world problems due to limited computational capacity. In the empirical experiments for black-box models, we show that the new analysis provides superior performance and faster convergence to a solution of nonconvex nonsmooth problems compared to the existing ZO-type methods as they suffer from small-level stepsizes. As a byproduct, the proposed analysis is generic and can be exploited to the other variants of gradient-free variance reduction methods aiming to make them more efficient.

Recently, data-driven inertial navigation approaches have demonstrated their capability of using well-trained neural networks to obtain accurate position estimates from inertial measurement units (IMUs) measurements. In this paper, we propose a novel robust Contextual Transformer-based network for Inertial Navigation (CTIN) to accurately predict velocity and trajectory. To this end, we first design a ResNet-based encoder enhanced by local and global multi-head self-attention to capture spatial contextual information from IMU measurements. Then we fuse these spatial representations with temporal knowledge by leveraging multi-head attention in the Transformer decoder. Finally, multi-task learning with uncertainty reduction is leveraged to improve learning efficiency and prediction accuracy of velocity and trajectory. Through extensive experiments over a wide range of inertial datasets (e.g., RIDI, OxIOD, RoNIN, IDOL, and our own), CTIN is very robust and outperforms state-of-the-art models.

Optimizing prediction accuracy can come at the expense of fairness. Towards minimizing discrimination against a group, fair machine learning algorithms strive to equalize the behavior of a model across different groups, by imposing a fairness constraint on models. However, we show that giving the same importance to groups of different sizes and distributions, to counteract the effect of bias in training data, can be in conflict with robustness. We analyze data poisoning attacks against group-based fair machine learning, with the focus on equalized odds. An adversary who can control sampling or labeling for a fraction of training data, can reduce the test accuracy significantly beyond what he can achieve on unconstrained models. Adversarial sampling and adversarial labeling attacks can also worsen the model's fairness gap on test data, even though the model satisfies the fairness constraint on training data. We analyze the robustness of fair machine learning through an empirical evaluation of attacks on multiple algorithms and benchmark datasets.

UK railway drainage systems are managed by a single infrastructure company called Network Rail (NR). These drainage systems are essential for conveying water away from the track and preventing operational and asset failures such as track surface flooding. With a detailed well calibrated hydraulic model of a drainage system, track flooding could be predicted for anticipated rainfall events, but the data does not exist to build accurate hydraulic models for the entire rail network (20,000 miles). Currently, drainage asset data is sparse, infrequently collected, and not organised into systems. In this study, machine learning (ML) algorithms were developed to analyse historical data of track flooding in order to identify plausible linkages between rainfall parameters, drainage asset condition, and track flooding (seen as failures), and to thus develop a predictive tool for drainage system failure causing track flooding. Plausible linkages between various input and output parameters were firstly identified and investigated in a qualitative and efficient way using an unsupervised ML technique, namely Self-Organising Maps (SOM). The plausibility of the identified linkages was further investigated by conducting interviews with experienced members of NR staff, and probable failure mechanisms were identified for track flooding, and presented in the form of failure pathways. Supervised ML algorithms were then employed to objectively quantify these linkages and estimate potential for track flooding failure under different rainfall scenarios using existing asset characterisation and condition data. The developed algorithms were then applied to understand the potential for failure risk for small groups of assets (1/8-mile sections of a UK railway line) for a range of rainfall values to examine under what condition any small system of assets can be expected to fail. Once trained, the model can be used at an operational level, where weather forecasts can be used to identify assets with immediate high risk, or at a strategic level, where all assets could be assessed for risk of failure under typical rainfall conditions, enabling risk ranking of assets for maintenance prioritisation. Ultimately this technique could provide a mechanism for improving drainage system management, reducing the risk of failure, and thus improving the safety and reliability of the UK's railway network.

We propose subsampling as a unified algorithmic technique for submodular maximization in centralized and online settings. The idea is simple: independently sample elements from the ground set, and use simple combinatorial techniques (such as greedy or local search) on these sampled elements. We show that this approach leads to optimal/state-of-the-art results despite being much simpler than existing methods. In the usual offline setting, we present SampleGreedy, which obtains a (p + 2 + o(1))-approximation for maximizing a submodular function subject to a p-extendible system using O(n + nk/p) evaluation and feasibility queries, where k is the size of the largest feasible set. The approximation ratio improves to p+1 and p for monotone submodular and linear objectives, respectively. In the streaming setting, we present SampleStreaming, which obtains a (4p +2 - o(1))-approximation for maximizing a submodular function subject to a p-matchoid using O(k) memory and O(km/p) evaluation and feasibility queries per element, where m is the number of matroids defining the p-matchoid. The approximation ratio improves to 4p for monotone submodular objectives. We empirically demonstrate the effectiveness of our algorithms on video summarization, location summarization, and movie recommendation tasks.

In this paper, we provide the first deterministic algorithm that achieves the tight $1-1/e$ approximation guarantee for submodular maximization under a cardinality (size) constraint while making a number of queries that scales only linearly with the size of the ground set $n$. To complement our result, we also show strong information-theoretic lower bounds. More specifically, we show that when the maximum cardinality allowed for a solution is constant, no algorithm making a sub-linear number of function evaluations can guarantee any constant approximation ratio. Furthermore, when the constraint allows the selection of a constant fraction of the ground set, we show that any algorithm making fewer than $\Omega(n/\log(n))$ function evaluations cannot perform better than an algorithm that simply outputs a uniformly random subset of the ground set of the right size. We then provide a variant of our deterministic algorithm for the more general knapsack constraint, which is the first linear-time algorithm that achieves $1/2$-approximation guarantee for this constraint. Finally, we extend our results to the general case of maximizing a monotone submodular function subject to the intersection of a $p$-set system and multiple knapsack constraints. We extensively evaluate the performance of our algorithms on multiple real-life machine learning applications, including movie recommendation, location summarization, twitter text summarization and video summarization.

White box adversarial perturbations are generated via iterative optimization algorithms most often by minimizing an adversarial loss on a ℓ _p neighborhood of the original image, the so-called distortion set. Constraining the adversarial search with different norms results in disparately structured adversarial examples. Here we explore several distortion sets with structure-enhancing algorithms. These new structures for adversarial examples might provide challenges for provable and empirical robust mechanisms. Because adversarial robustness is still an empirical field, defense mechanisms should also reasonably be evaluated against differently structured attacks. Besides, these structured adversarial perturbations may allow for larger distortions size than their ℓ _p counterpart while remaining imperceptible or perceptible as natural distortions of the image. We will demonstrate in this work that the proposed structured adversarial examples can significantly bring down the classification accuracy of adversarially trained classifiers while showing a low ℓ _2 distortion rate. For instance, on ImagNet dataset the structured attacks drop the accuracy of the adversarial model to near zero with only 50 ℓ _2 distortion generated using white-box attacks like PGD. As a byproduct, our findings on structured adversarial examples can be used for adversarial regularization of models to make models more robust or improve their generalization performance on datasets that are structurally different.

White box adversarial perturbations are generated via iterative optimization algorithms most often by minimizing an adversarial loss on a $\ell_p$ neighborhood of the original image, the so-called distortion set. Constraining the adversarial search with different norms results in disparately structured adversarial examples. Here we explore several distortion sets with structure-enhancing algorithms. These new structures for adversarial examples might provide challenges for provable and empirical robust mechanisms. Because adversarial robustness is still an empirical field, defense mechanisms should also reasonably be evaluated against differently structured attacks. Besides, these structured adversarial perturbations may allow for larger distortions size than their $\ell_p$ counter-part while remaining imperceptible or perceptible as natural distortions of the image. We will demonstrate in this work that the proposed structured adversarial examples can significantly bring down the classification accuracy of adversarialy trained classifiers while showing low $\ell_2$ distortion rate. For instance, on ImagNet dataset the structured attacks drop the accuracy of adversarial model to near zero with only 50\% of $\ell_2$ distortion generated using white-box attacks like PGD. As a byproduct, our finding on structured adversarial examples can be used for adversarial regularization of models to make models more robust or improve their generalization performance on datasets which are structurally different.

We propose an ensemble-based defense against adversarial examples using distance map layers (DMLs). Similar to fully connected layers, DMLs can be used to output logits for a multi-class classification model. We show in this paper how DMLs can be deployed to prevent transferability of attacks across ensemble members by adapting pairwise (almost) orthogonal covariance matrices. We also illustrate how DMLs provide an efficient way to regularize the Lipschitz constant of the ensemble's member models, which further boosts the resulting robustness. Through empirical evaluations across multiple datasets and attack models, we demonstrate that the ensembles based on DMLs can achieve high benign accuracy while exhibiting robustness against adversarial attacks using multiple white-box techniques along with AutoAttack.

In this paper, we propose scalable methods for maximizing a regularized submodular function $f = g - \ell$ expressed as the difference between a monotone submodular function $g$ and a modular function $\ell$. Indeed, submodularity is inherently related to the notions of diversity, coverage, and representativeness. In particular, finding the mode of many popular probabilistic models of diversity, such as determinantal point processes, submodular probabilistic models, and strongly log-concave distributions, involves maximization of (regularized) submodular functions. Since a regularized function $f$ can potentially take on negative values, the classic theory of submodular maximization, which heavily relies on the non-negativity assumption of submodular functions, may not be applicable. To circumvent this challenge, we develop the first one-pass streaming algorithm for maximizing a regularized submodular function subject to a $k$-cardinality constraint. It returns a solution $S$ with the guarantee that $f(S)\geq(\phi^{-2}-\epsilon) \cdot g(OPT)-\ell (OPT)$, where $\phi$ is the golden ratio. Furthermore, we develop the first distributed algorithm that returns a solution $S$ with the guarantee that $\mathbb{E}[f(S)] \geq (1-\epsilon) [(1-e^{-1}) \cdot g(OPT)-\ell(OPT)]$ in $O(1/ \epsilon)$ rounds of MapReduce computation, without keeping multiple copies of the entire dataset in each round (as it is usually done). We should highlight that our result, even for the unregularized case where the modular term $\ell$ is zero, improves the memory and communication complexity of the existing work by a factor of $O(1/ \epsilon)$ while arguably provides a simpler distributed algorithm and a unifying analysis. We also empirically study the performance of our scalable methods on a set of real-life applications, including finding the mode of distributions, data summarization, and product recommendation.

In this paper, we introduce a novel technique for constrained submodular maximization, inspired by barrier functions in continuous optimization. This connection not only improves the running time for constrained submodular maximization but also provides the state of the art guarantee. More precisely, for maximizing a monotone submodular function subject to the combination of a $k$-matchoid and $\ell$-knapsack constraint (for $\ell\leq k$), we propose a potential function that can be approximately minimized. Once we minimize the potential function up to an $\epsilon$ error it is guaranteed that we have found a feasible set with a $2(k+1+\epsilon)$-approximation factor which can indeed be further improved to $(k+1+\epsilon)$ by an enumeration technique. We extensively evaluate the performance of our proposed algorithm over several real-world applications, including a movie recommendation system, summarization tasks for YouTube videos, Twitter feeds and Yelp business locations, and a set cover problem.

Protein-protein interaction (PPI) network alignment is a canonical operation to transfer biological knowledge among species. The alignment of PPI-networks has many applications, such as the prediction of protein function, detection of conserved network motifs, and the reconstruction of species' phylogenetic relationships. A good multiple-network alignment (MNA), by considering the data related to several species, provides a deep understanding of biological networks and system-level cellular processes. With the massive amounts of available PPI data and the increasing number of known PPI networks, the problem of MNA is gaining more attention in the systems-biology studies. In this paper, we introduce a new scalable and accurate algorithm, called MPGM, for aligning multiple networks. The MPGM algorithm has two main steps: (i) SEEDGENERATION and (ii) MULTIPLEPERCOLATION. In the first step, to generate an initial set of seed tuples, the SEEDGENERATION algorithm uses only protein sequence similarities. In the second step, to align remaining unmatched nodes, the MULTIPLEPERCOLATION algorithm uses network structures and the seed tuples generated from the first step. We show that, with respect to different evaluation criteria, MPGM outperforms the other state-of-the-art algorithms. In addition, we guarantee the performance of MPGM under certain classes of network models. We introduce a sampling-based stochastic model for generating k correlated networks. We prove that for this model, if a sufficient number of seed tuples are available, the MULTIPLEPERCOLATION algorithm correctly aligns almost all the nodes. Our theoretical results are supported by experimental evaluations over synthetic networks.

In this paper, we propose a novel framework that converts streaming algorithms for monotone submodular maximization into streaming algorithms for non-monotone submodular maximization. This reduction readily leads to the currently tightest deterministic approximation ratio for submodular maximization subject to a $k$-matchoid constraint. Moreover, we propose the first streaming algorithm for monotone submodular maximization subject to $k$-extendible and $k$-set system constraints. Together with our proposed reduction, we obtain $O(k\log k)$ and $O(k^2\log k)$ approximation ratio for submodular maximization subject to the above constraints, respectively. We extensively evaluate the empirical performance of our algorithm against the existing work in a series of experiments including finding the maximum independent set in randomly generated graphs, maximizing linear functions over social networks, movie recommendation, Yelp location summarization, and Twitter data summarization.

Nektar++ is an open-source framework that provides a flexible, high-performance and scalable platform for the development of solvers for partial differential equations using the high-order spectral/$hp$ element method. In particular, Nektar++ aims to overcome the complex implementation challenges that are often associated with high-order methods, thereby allowing them to be more readily used in a wide range of application areas. In this paper, we present the algorithmic, implementation and application developments associated with our Nektar++ version 5.0 release. We describe some of the key software and performance developments, including our strategies on parallel I/O, on in situ processing, the use of collective operations for exploiting current and emerging hardware, and interfaces to enable multi-solver coupling. Furthermore, we provide details on a newly developed Python interface that enables a more rapid introduction for new users unfamiliar with spectral/$hp$ element methods, C++ and/or Nektar++. This release also incorporates a number of numerical method developments - in particular: the method of moving frames, which provides an additional approach for the simulation of equations on embedded curvilinear manifolds and domains; a means of handling spatially variable polynomial order; and a novel technique for quasi-3D simulations to permit spatially-varying perturbations to the geometry in the homogeneous direction. Finally, we demonstrate the new application-level features provided in this release, namely: a facility for generating high-order curvilinear meshes called NekMesh; a novel new AcousticSolver for aeroacoustic problems; our development of a 'thick' strip model for the modelling of fluid-structure interaction problems in the context of vortex-induced vibrations. We conclude by commenting some directions for future code development and expansion.