Attention models are typically learned by optimizing one of three standard loss functions that are variously called -- soft attention, hard attention, and latent variable marginal likelihood (LVML) attention. All three paradigms are motivated by the same goal of finding two models -- a `focus' model that `selects' the right \textit{segment} of the input and a `classification' model that processes the selected segment into the target label. However, they differ significantly in the way the selected segments are aggregated, resulting in distinct dynamics and final results. We observe a unique signature of models learned using these paradigms and explain this as a consequence of the evolution of the classification model under gradient descent when the focus model is fixed. We also analyze these paradigms in a simple setting and derive closed-form expressions for the parameter trajectory under gradient flow. With the soft attention loss, the focus model improves quickly at initialization and splutters later on. On the other hand, hard attention loss behaves in the opposite fashion. Based on our observations, we propose a simple hybrid approach that combines the advantages of the different loss functions and demonstrates it on a collection of semi-synthetic and real-world datasets
相關內容
We study the forgetting properties of the particle filter when its state - the collection of particles - is regarded as a Markov chain. Under a strong mixing assumption on the particle filter's underlying Feynman-Kac model, we find that the particle filter is exponentially mixing, and forgets its initial state in $O(\log N )$ `time', where $N$ is the number of particles and time refers to the number of particle filter algorithm steps, each comprising a selection (or resampling) and mutation (or prediction) operation. We present an example which suggests that this rate is optimal. In contrast to our result, available results to-date are extremely conservative, suggesting $O(\alpha^N)$ time steps are needed, for some $\alpha>1$, for the particle filter to forget its initialisation. We also study the conditional particle filter (CPF) and extend our forgetting result to this context. We establish a similar conclusion, namely, CPF is exponentially mixing and forgets its initial state in $O(\log N )$ time. To support this analysis, we establish new time-uniform $L^p$ error estimates for CPF, which can be of independent interest.
Soft pneumatic actuators are used to steer soft growing "vine" robots while being flexible enough to undergo the tip eversion required for growth. In this study, we compared the performance of three types of pneumatic actuators in terms of their ability to perform eversion, quasi-static bending, dynamic motion, and force output: the pouch motor, the cylindrical pneumatic artificial muscle (cPAM), and the fabric pneumatic artificial muscle (fPAM). The pouch motor is advantageous for prototyping due to its simple manufacturing process. The cPAM exhibits superior bending behavior and produces the highest forces, while the fPAM actuates fastest and everts at the lowest pressure. We evaluated a range of dimensions for each actuator type. Larger actuators can produce more significant deformations and forces, but smaller actuators inflate faster and can evert at a lower pressure. Because vine robots are lightweight, the effect of gravity on the functionality of different actuators is minimal. We developed a new analytical model that predicts the pressure-to-bending behavior of vine robot actuators. Using the actuator results, we designed and demonstrated a 4.8 m long vine robot equipped with highly maneuverable 60x60 mm cPAMs in a three-dimensional obstacle course. The vine robot was able to move around sharp turns, travel through a passage smaller than its diameter, and lift itself against gravity.
Software debloating techniques are applied to craft a specialized version of the program based on the user's requirements and remove irrelevant code accordingly. The debloated programs presumably maintain better performance and reduce the attack surface in contrast to the original programs. This work unleashes the effectiveness of applying software debloating techniques on the robustness of machine learning systems in the malware classification domain. We empirically study how an adversarial can leverage software debloating techniques to mislead machine learning malware classification models. We apply software debloating techniques to generate adversarial examples and demonstrate these adversarial examples can reduce the detection rate of VirusTotal. Our study opens new directions for research into adversarial machine learning not only in malware detection/classification but also in other software domains.
Traditional approaches for manipulation planning rely on an explicit geometric model of the environment to formulate a given task as an optimization problem. However, inferring an accurate model from raw sensor input is a hard problem in itself, in particular for articulated objects (e.g., closets, drawers). In this paper, we propose a Neural Field Representation (NFR) of articulated objects that enables manipulation planning directly from images. Specifically, after taking a few pictures of a new articulated object, we can forward simulate its possible movements, and, therefore, use this neural model directly for planning with trajectory optimization. Additionally, this representation can be used for shape reconstruction, semantic segmentation and image rendering, which provides a strong supervision signal during training and generalization. We show that our model, which was trained only on synthetic images, is able to extract a meaningful representation for unseen objects of the same class, both in simulation and with real images. Furthermore, we demonstrate that the representation enables robotic manipulation of an articulated object in the real world directly from images.
This paper investigates the performance of a singleuser fluid antenna system (FAS), by exploiting a class of elliptical copulas to describe the structure of dependency amongst the fluid antenna ports. By expressing Jakes' model in terms of the Gaussian copula, we consider two cases: (i) the general case, i.e., any arbitrary correlated fading distribution; and (ii) the specific case, i.e., correlated Nakagami-m fading. For both scenarios, we first derive analytical expressions for the cumulative distribution function (CDF) and probability density function (PDF) of the equivalent channel in terms of multivariate normal distribution. Then, we obtain the outage probability (OP) and the delay outage rate (DOR) to analyze the performance of the FAS. By employing the popular rank correlation coefficients such as Spearman's \{rho} and Kendall's {\tau}, we measure the degree of dependency in correlated arbitrary fading channels and illustrate how the Gaussian copula can be accurately connected to Jakes' model in FAS without complicated mathematical analysis. Numerical results show that increasing the fluid antenna size provides lower OP and DOR, but the system performance saturates as the number of antenna ports increases. In addition, our results indicate that FAS provides better performance compared to conventional single-fixed antenna systems even when the size of fluid antenna is small.
Diffusion models have demonstrated impressive generative capabilities, but their 'exposure bias' problem, described as the input mismatch between training and sampling, lacks in-depth exploration. In this paper, we systematically investigate the exposure bias problem in diffusion models by first analytically modelling the sampling distribution, based on which we then attribute the prediction error at each sampling step as the root cause of the exposure bias issue. Furthermore, we discuss potential solutions to this issue and propose an intuitive metric for it. Along with the elucidation of exposure bias, we propose a simple, yet effective, training-free method called Epsilon Scaling to alleviate the exposure bias. We show that Epsilon Scaling explicitly moves the sampling trajectory closer to the vector field learned in the training phase by scaling down the network output (Epsilon), mitigating the input mismatch between training and sampling. Experiments on various diffusion frameworks (ADM, DDPM/DDIM, EDM, LDM), unconditional and conditional settings, and deterministic vs. stochastic sampling verify the effectiveness of our method. The code is available at //github.com/forever208/ADM-ES; //github.com/forever208/EDM-ES
Adversarial attacks expose vulnerabilities of deep learning models by introducing minor perturbations to the input, which lead to substantial alterations in the output. Our research focuses on the impact of such adversarial attacks on sequence-to-sequence (seq2seq) models, specifically machine translation models. We introduce algorithms that incorporate basic text perturbation heuristics and more advanced strategies, such as the gradient-based attack, which utilizes a differentiable approximation of the inherently non-differentiable translation metric. Through our investigation, we provide evidence that machine translation models display robustness displayed robustness against best performed known adversarial attacks, as the degree of perturbation in the output is directly proportional to the perturbation in the input. However, among underdogs, our attacks outperform alternatives, providing the best relative performance. Another strong candidate is an attack based on mixing of individual characters.
As artificial intelligence (AI) models continue to scale up, they are becoming more capable and integrated into various forms of decision-making systems. For models involved in moral decision-making, also known as artificial moral agents (AMA), interpretability provides a way to trust and understand the agent's internal reasoning mechanisms for effective use and error correction. In this paper, we provide an overview of this rapidly-evolving sub-field of AI interpretability, introduce the concept of the Minimum Level of Interpretability (MLI) and recommend an MLI for various types of agents, to aid their safe deployment in real-world settings.
AI is undergoing a paradigm shift with the rise of models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks. We call these models foundation models to underscore their critically central yet incomplete character. This report provides a thorough account of the opportunities and risks of foundation models, ranging from their capabilities (e.g., language, vision, robotics, reasoning, human interaction) and technical principles(e.g., model architectures, training procedures, data, systems, security, evaluation, theory) to their applications (e.g., law, healthcare, education) and societal impact (e.g., inequity, misuse, economic and environmental impact, legal and ethical considerations). Though foundation models are based on standard deep learning and transfer learning, their scale results in new emergent capabilities,and their effectiveness across so many tasks incentivizes homogenization. Homogenization provides powerful leverage but demands caution, as the defects of the foundation model are inherited by all the adapted models downstream. Despite the impending widespread deployment of foundation models, we currently lack a clear understanding of how they work, when they fail, and what they are even capable of due to their emergent properties. To tackle these questions, we believe much of the critical research on foundation models will require deep interdisciplinary collaboration commensurate with their fundamentally sociotechnical nature.
Object detection typically assumes that training and test data are drawn from an identical distribution, which, however, does not always hold in practice. Such a distribution mismatch will lead to a significant performance drop. In this work, we aim to improve the cross-domain robustness of object detection. We tackle the domain shift on two levels: 1) the image-level shift, such as image style, illumination, etc, and 2) the instance-level shift, such as object appearance, size, etc. We build our approach based on the recent state-of-the-art Faster R-CNN model, and design two domain adaptation components, on image level and instance level, to reduce the domain discrepancy. The two domain adaptation components are based on H-divergence theory, and are implemented by learning a domain classifier in adversarial training manner. The domain classifiers on different levels are further reinforced with a consistency regularization to learn a domain-invariant region proposal network (RPN) in the Faster R-CNN model. We evaluate our newly proposed approach using multiple datasets including Cityscapes, KITTI, SIM10K, etc. The results demonstrate the effectiveness of our proposed approach for robust object detection in various domain shift scenarios.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191