Multi-Task Learning (MTL) involves the concurrent training of multiple tasks, offering notable advantages for dense prediction tasks in computer vision. MTL not only reduces training and inference time as opposed to having multiple single-task models, but also enhances task accuracy through the interaction of multiple tasks. However, existing methods face limitations. They often rely on suboptimal cross-task interactions, resulting in task-specific predictions with poor geometric and predictive coherence. In addition, many approaches use inadequate loss weighting strategies, which do not address the inherent variability in task evolution during training. To overcome these challenges, we propose an advanced MTL model specifically designed for dense vision tasks. Our model leverages state-of-the-art vision transformers with task-specific decoders. To enhance cross-task coherence, we introduce a trace-back method that improves both cross-task geometric and predictive features. Furthermore, we present a novel dynamic task balancing approach that projects task losses onto a common scale and prioritizes more challenging tasks during training. Extensive experiments demonstrate the superiority of our method, establishing new state-of-the-art performance across two benchmark datasets. The code is available at: https://github.com/Klodivio355/MT-CP

Abstract Deep neural networks are vulnerable to adversarial examples. Adversarial training (AT) is an effective defense against adversarial examples. However, AT is prone to overfitting which degrades robustness substantially. Recently, data augmentation (DA) was shown to be effective in mitigating robust overfitting if appropriately designed and optimized for AT. This work proposes a new method to automatically learn online, instance-wise, DA policies to improve robust generalization for AT. This is the first automated DA method specific for robustness. A novel policy learning objective, consisting of Vulnerability, Affinity and Diversity, is proposed and shown to be sufficiently effective and efficient to be practical for automatic DA generation during AT. Importantly, our method dramatically reduces the cost of policy search from the 5000 h of AutoAugment and the 412 h of IDBH to 9 h, making automated DA more practical to use for adversarial robustness. This allows our method to efficiently explore a large search space for a more effective DA policy and evolve the policy as training progresses. Empirically, our method is shown to outperform all competitive DA methods across various model architectures and datasets. Our DA policy reinforced vanilla AT to surpass several state-of-the-art AT methods regarding both accuracy and robustness. It can also be combined with those advanced AT methods to further boost robustness. Code and pre-trained models are available at: https://github.com/TreeLLi/AROID .

Reliable and robust evaluation methods are a necessary first step towards developing machine learning models that are themselves robust and reliable. Unfortunately, current evaluation protocols typically used to assess classifiers fail to comprehensively evaluate performance as they tend to rely on limited types of test data, and ignore others. For example, using the standard test data fails to evaluate the predictions made by the classifier to samples from classes it was not trained on. On the other hand, testing with data containing samples from unknown classes fails to evaluate how well the classifier can predict the labels for known classes. This article advocates benchmarking performance using a wide range of different types of data and using a single metric that can be applied to all such data types to produce a consistent evaluation of performance. Using the proposed benchmark it is found that current deep neural networks, including those trained with methods that are believed to produce state-of-the-art robustness, are vulnerable to making mistakes on certain types of data. This means that such models will be unreliable in real-world scenarios where they may encounter data from many different domains, and that they are insecure as they can be easily fooled into making the wrong decisions. It is hoped that these results will motivate the wider adoption of more comprehensive testing methods that will, in turn, lead to the development of more robust machine learning methods in the future. Code: https://codeberg.org/mwspratling/RobustnessEvaluation.

Deep neural networks are highly susceptible to adversarial perturbations: artificial noise that corrupts input data in ways imperceptible to humans but causes incorrect predictions. Among the various defenses against these attacks, adversarial training has emerged as the most effective. In this work, we aim to enhance adversarial training to improve robustness against adversarial attacks. We begin by analyzing how adversarial vulnerability evolves during training from an instance-wise perspective. This analysis reveals two previously unrecognized phenomena: robust shortcut and disordered robustness . We then demonstrate that these phenomena are related to robust overfitting , a well-known issue in adversarial training. Building on these insights, we propose a novel adversarial training method: Instance-adaptive Smoothness Enhanced Adversarial Training (ISEAT). This method jointly smooths the input and weight loss landscapes in an instance-adaptive manner, preventing the exploitation of robust shortcut and thereby mitigating robust overfitting. Extensive experiments demonstrate the efficacy of ISEAT and its superiority over existing adversarial training methods. Code is available at https://github.com/TreeLLi/ISEAT . • Adversarial training can overfit through robust shortcut. • Some training data exhibit disordered robustness after adversarial training. • Adversarial training can be improved via instance adaptive loss smoothing.

Most existing anomaly detection (AD) methods require a dedicated model for each category. Such a paradigm, despite its promising results, is computationally expensive and inefficient, thereby failing to meet the requirements for real-world applications. Inspired by how humans detect anomalies, by comparing a query image to known normal ones, this article proposes a novel few-shot AD (FSAD) framework. Using a training set of normal images from various categories, registration, aiming to align normal images of the same categories, is leveraged as the proxy task for self-supervised category-agnostic representation learning. At test time, an image and its corresponding support set, consisting of a few normal images from the same category, are supplied, and anomalies are identified by comparing the registered features of the test image to its corresponding support image features. Such a setup enables the model to generalize to novel test categories. It is, to our best knowledge, the first FSAD method that requires no model fine-tuning for novel categories: enabling a single model to be applied to all categories. Extensive experiments demonstrate the effectiveness of the proposed method. Particularly, it improves the current state-of-the-art (SOTA) for FSAD by 11.3% and 8.3% on the MVTec and MPDD benchmarks, respectively. The source code is available at https://github.com/Haoyan-Guan/CAReg.

Large pre-trained Vision-Language Models (VLMs) like CLIP, despite having remarkable generalization ability, are highly vulnerable to adversarial examples. This work studies the adversarial robustness of VLMs from the novel perspective of the text prompt instead of the extensively studied model weights (frozen in this work). We first show that the effectiveness of both adversarial attack and defense are sensitive to the used text prompt. Inspired by this, we propose a method to improve resilience to adversarial attacks by learning a robust text prompt for VLMs. The proposed method, named Adversarial Prompt Tuning (APT), is effective while being both computationally and data efficient. Extensive experiments are conducted across 15 datasets and 4 data sparsity schemes (from 1-shot to full training data settings) to show APT's superiority over hand-engineered prompts and other state-of-the-art adaption methods. APT demonstrated excellent abilities in terms of the in-distribution performance and the generalization under input distribution shift and across datasets. Surprisingly, by simply adding one learned word to the prompts, APT can significantly boost the accuracy and robustness ($\epsilon=4/255$) over the hand-engineered prompts by +13% and +8.5% on average respectively. The improvement further increases, in our most effective setting, to +26.4% for accuracy and +16.7% for robustness. Code is available at https://github.com/TreeLLi/APT.

Existing works have made great progress in improving adversarial robustness, but typically test their method only on data from the same distribution as the training data, i.e. in-distribution (ID) testing. As a result, it is unclear how such robustness generalizes under input distribution shifts, i.e. out-of-distribution (OOD) testing. This omission is concerning as such distribution shifts are unavoidable when methods are deployed in the wild. To address this issue we propose a benchmark named OODRobustBench to comprehensively assess OOD adversarial robustness using 23 dataset-wise shifts (i.e. naturalistic shifts in input distribution) and 6 threat-wise shifts (i.e., unforeseen adversarial threat models). OODRobustBench is used to assess 706 robust models using 60.7K adversarial evaluations. This large-scale analysis shows that: 1) adversarial robustness suffers from a severe OOD generalization issue; 2) ID robustness correlates strongly with OOD robustness in a positive linear way. The latter enables the prediction of OOD robustness from ID robustness. We then predict and verify that existing methods are unlikely to achieve high OOD robustness. Novel methods are therefore required to achieve OOD robustness beyond our prediction. To facilitate the development of these methods, we investigate a wide range of techniques and identify several promising directions. Code and models are available at: https://github.com/OODRobustBench/OODRobustBench.

Abstract Modelling one‐to‐many type mappings in problems with a temporal component can be challenging. Backpropagation is not applicable to networks that perform discrete sampling and is also susceptible to gradient instabilities, especially when applied to longer sequences. In this paper, we propose two recurrent neural network architectures that leverage stochastic units and mixture models, and are trained with target propagation. We demonstrate that these networks can model complex conditional probability distributions, outperform backpropagation‐trained alternatives, and do not rapidly degrade with increased time horizons. Our main contributions consist of the design and evaluation of the architectures that enable the networks to solve multi‐model problems with a temporal dimension. This also includes the extension of the target propagation through time algorithm to handle stochastic neurons. The use of target propagation provides an additional computational advantage, which enables the network to handle time horizons that are substantially longer compared to networks fitted using backpropagation.

Multitask learning (MTL) aims to learn multiple tasks simultaneously while exploiting their mutual relationships. By using shared resources to simultaneously calculate multiple outputs, this learning paradigm has the potential to have lower memory requirements and inference times compared to the traditional approach of using separate methods for each task. Previous work in MTL has mainly focused on fully supervised methods, as task relationships (TRs) can not only be leveraged to lower the level of data dependency of those methods but also improve the performance. However, MTL introduces a set of challenges due to a complex optimization scheme and a higher labeling requirement. This article focuses on how MTL could be utilized under different partial supervision settings to address these challenges. First, this article analyses how MTL traditionally uses different parameter sharing techniques to transfer knowledge in between tasks. Second, it presents different challenges arising from such a multiobjective optimization (MOO) scheme. Third, it introduces how task groupings (TGs) can be achieved by analyzing TRs. Fourth, it focuses on how partially supervised methods applied to MTL can tackle the aforementioned challenges. Lastly, this article presents the available datasets, tools, and benchmarking results of such methods. The reviewed articles, categorized following this work, are available athttps://github.com/Klodivio355/MTL-CV-Review.

Abstract Few-shot segmentation (FSS) aims to segment unseen classes using a few annotated samples. Typically, a prototype representing the foreground class is extracted from annotated support image(s) and is matched to features representing each pixel in the query image. However, models learnt in this way are insufficiently discriminatory, and often produce false positives: misclassifying background pixels as foreground. Some FSS methods try to address this issue by using the background in the support image(s) to help identify the background in the query image. However, the backgrounds of these images are often quite distinct, and hence, the support image background information is uninformative. This article proposes a method, QSR, that extracts the background from the query image itself, and as a result is better able to discriminate between foreground and background features in the query image. This is achieved by modifying the training process to associate prototypes with class labels including known classes from the training data and latent classes representing unknown background objects. This class information is then used to extract a background prototype from the query image. To successfully associate prototypes with class labels and extract a background prototype that is capable of predicting a mask for the background regions of the image, the machinery for extracting and using foreground prototypes is induced to become more discriminative between different classes. Experiments achieves state-of-the-art results for both 1-shot and 5-shot FSS on the PASCAL- $$5^{i}$$ 5 i and COCO- $$20^{i}$$ 20 i dataset. As QSR operates only during training, results are produced with no extra computational complexity during testing.

Adversarial training is widely used to improve the robustness of deep neural networks to adversarial attack. However, adversarial training is prone to overfitting, and the cause is far from clear. This work sheds light on the mechanisms underlying overfitting through analyzing the loss landscape w.r.t. the input. We find that robust overfitting results from standard training, specifically the minimization of the clean loss, and can be mitigated by regularization of the loss gradients. Moreover, we find that robust overfitting turns severer during adversarial training partially because the gradient regularization effect of adversarial training becomes weaker due to the increase in the loss landscape’s curvature. To improve robust generalization, we propose a new regularizer to smooth the loss landscape by penalizing the weighted logits variation along the adversarial direction. Our method significantly mitigates robust overfitting and achieves the highest robustness and efficiency compared to similar previous methods. Code is available at https://github.com/TreeLLi/Combating-RO-AdvLC.

Abstract Many current trackers utilise an appearance model to localise the target object in each frame. However, such approaches often fail when there are similar-looking distractor objects in the surrounding background, meaning that target appearance alone is insufficient for robust tracking. In contrast, humans consider the distractor objects as additional visual cues, in order to infer the position of the target. Inspired by this observation, this paper proposes a novel tracking architecture in which not only is the appearance of the tracked object, but also the appearance of the distractors detected in previous frames, taken into consideration using a form of probabilistic inference known as explaining away. This mechanism increases the robustness of tracking by making it more likely that the target appearance model is matched to the true target, rather than similar-looking regions of the current frame. The proposed method can be combined with many existing trackers. Combining it with SiamFC, DaSiamRPN, Super_DiMP, and ARSuper_DiMP all resulted in an increase in the tracking accuracy compared to that achieved by the underlying tracker alone. When combined with Super_DiMP and ARSuper_DiMP, the resulting trackers produce performance that is competitive with the state of the art on seven popular benchmarks.

Few-shot segmentation aims to segment images containing objects from previously unseen classes using only a few annotated samples. Most current methods focus on using object information extracted, with the aid of human annotations, from support images to identify the same objects in new query images. However, background information can also be useful to distinguish objects from their surroundings. Hence, some previous methods also extract background information from the support images. In this paper, we argue that such information is of limited utility, as the background in different images can vary widely. To overcome this issue, we propose CobNet which utilises information about the background that is extracted from the query images without annotations of those images. Experiments show that our method achieves a mean Intersection-over-Union score of 61.4% and 37.8% for 1-shot segmentation on PASCAL-5iand COCO-20irespectively, outperforming previous methods. It is also shown to produce state-of-the-art performances of 53.7% for weakly-supervised few-shot segmentation, where no annotations are provided for the support images.

Most current trackers utilise an appearance model to localise the target object in each frame. However, such approaches often fail when there are similar looking distractor objects in the surrounding background. This paper promotes an approach that can be combined with many existing trackers to tackle this issue and improve tracking robustness. The proposed approach makes use of two additional cues to target location: shape cues which are exploited through offline training of the appearance model, and motion cues which are exploited online to predict the target object’s future position based on its history of past locations. Combining these additional mechanisms with the existing trackers SiamFC, SiamFC++, Super_DiMP and ARSuper_DiMP all resulted in an increase in the tracking accuracy compared to that achieved by the corresponding underlying tracker alone. When combined with ARSuper_DiMP the resulting tracker is shown to outperform all popular state-of-the-art trackers on three benchmark datasets (OTB-100, NFS, and LaSOT), and produce performance that is competitive with the state-of-the-art on the UAV123, Trackingnet, GOT-10K and VOT2020 datasets.

Recurrent Neural Networks have been widely used to process sequence data, but have long been criticized for their biological implausibility and training difficulties related to vanishing and exploding gradients. This paper presents a novel algorithm for training recurrent networks, target propagation through time (TPTT), that outperforms standard backpropagation through time (BPTT) on four out of the five problems used for testing. The proposed algorithm is initially tested and compared to BPTT on four synthetic time lag tasks, and its performance is also measured using the sequential MNIST data set. In addition, as TPTT uses target propagation, it allows for discrete nonlinearities and could potentially mitigate the credit assignment problem in more complex recurrent architectures.

Predictive coding (PC) is a leading theory of cortical function that has previously been shown to explain a great deal of neurophysiological and psychophysical data. Here it is shown that PC can perform almost exact Bayesian inference when applied to computing with population codes. It is demonstrated that the proposed algorithm, based on PC, can: decode probability distributions encoded as noisy population codes; combine priors with likelihoods to calculate posteriors; perform cue integration and cue segregation; perform function approximation; be extended to perform hierarchical inference; simultaneously represent and reason about multiple stimuli; and perform inference with multi-modal and non-Gaussian probability distributions. PC thus provides a neural network-based method for performing probabilistic computation and provides a simple, yet comprehensive, theory of how the cerebral cortex performs Bayesian inference.

Algorithms that encode images using a sparse set of basis functions have previously been shown to explain aspects of the physiology of a primary visual cortex (V1), and have been used for applications, such as image compression, restoration, and classification. Here, a sparse coding algorithm, that has previously been used to account for the response properties of orientation tuned cells in primary visual cortex, is applied to the task of perceptually salient boundary detection. The proposed algorithm is currently limited to using only intensity information at a single scale. However, it is shown to out-perform the current state-of-the-art image segmentation method (Pb) when this method is also restricted to using the same information.

A method is presented for learning the reciprocal feedforward and feedback connections required by the predictive coding model of cortical function. When this method is used, feedforward and feedback connections are learned simultaneously and independently in a biologically plausible manner. The performance of the proposed algorithm is evaluated by applying it to learning the elementary components of artificial and natural images. For artificial images, the bars problem is employed, and the proposed algorithm is shown to produce state-of-the-art performance on this task. For natural images, components resembling Gabor functions are learned in the first processing stage, and neurons responsive to corners are learned in the second processing stage. The properties of these learned representations are in good agreement with neurophysiological data from V1 and V2. The proposed algorithm demonstrates for the first time that a single computational theory can explain the formation of cortical RFs and also the response properties of cortical neurons once those RFs have been learned.

The combination of two or more population-coded signals in a neural model of predictive coding can give rise to multiplicative gain modulation in the response properties of individual neurons. Synaptic weights generating these multiplicative response properties can be learned using an unsupervised, Hebbian learning rule. The behavior of the model is compared to empirical data on gaze-dependent gain modulation of cortical cells and found to be in good agreement with a range of physiological observations. Furthermore, it is demonstrated that the model can learn to represent a set of basis functions. This letter thus connects an often-observed neurophysiological phenomenon and important neurocomputational principle (gain modulation) with an influential theory of brain operation (predictive coding).

In order to perform object recognition it is necessary to learn representations of the underlying components of images. Such components correspond to objects, object-parts, or features. Non-negative matrix factorisation is a generative model that has been specifically proposed for finding such meaningful representations of image data, through the use of non-negativity constraints on the factors. This article reports on an empirical investigation of the performance of non-negative matrix factorisation algorithms. It is found that such algorithms need to impose additional constraints on the sparseness of the factors in order to successfully deal with occlusion. However, these constraints can themselves result in these algorithms failing to identify image components under certain conditions. In contrast, a recognition model (a competitive learning neural network algorithm) reliably and accurately learns representations of elementary image features without such constraints.

In order to perform object recognition, it is necessary to form perceptual representations that are sufficiently specific to distinguish between objects, but that are also sufficiently flexible to generalize across changes in location, rotation, and scale. A standard method for learning perceptual representations that are invariant to viewpoint is to form temporal associations across image sequences showing object transformations. However, this method requires that individual stimuli be presented in isolation and is therefore unlikely to succeed in real-world applications where multiple objects can co-occur in the visual input. This paper proposes a simple modification to the learning method that can overcome this limitation and results in more robust learning of invariant representations.

A large and influential class of neural network architectures uses postintegration lateral inhibition as a mechanism for competition. We argue that these algorithms are computationally deficient in that they fail to generate, or learn, appropriate perceptual representations under certain circumstances. An alternative neural network architecture is presented here in which nodes compete for the right to receive inputs rather than for the right to generate outputs. This form of competition, implemented through preintegration lateral inhibition, does provide appropriate coding properties and can be used to learn such representations efficiently. Furthermore, this architecture is consistent with both neuroanatomical and neurophysiological data. We thus argue that preintegration lateral inhibition has computational advantages over conventional neural network architectures while remaining equally biologically plausible.

Visual servoing is a process to enable a robot to position a camera with respect to known landmarks using the visual data obtained by the camera itself to guide camera motion. A solution is described which requires very little a priori information freeing it from being specific to a particular configuration of robot and camera. The solution is based on closed loop control together with deliberate perturbations of the trajectory to provide calibration movements for refining that trajectory. Results from experiments in simulation and on a physical robot arm (camera-in-hand configuration) are presented. 1 Introduction Visual servoing is a process by which the appearance of landmarks is used to control the positioning of the camera with respect to the world. The camera is thus the sensor for a control scheme, in which the position of the camera itself is the object of control. The objective is for a robot to position the camera in a specific `target' pose (defined at initialisati...