Despite deep learning (DL) has achieved remarkable progress in various domains, the DL models are still prone to making mistakes. This issue necessitates effective debugging tools for DL practitioners to interpret the decision-making process within the networks. However, existing debugging methods often demand extra data or adjustments to the decision process, limiting their applicability. To tackle this problem, we present NeuroInspect, an interpretable neuron-based debugging framework with three key stages: counterfactual explanations, feature visualizations, and false correlation mitigation. Our debugging framework first pinpoints neurons responsible for mistakes in the network and then visualizes features embedded in the neurons to be human-interpretable. To provide these explanations, we introduce CLIP-Illusion, a novel feature visualization method that generates images representing features conditioned on classes to examine the connection between neurons and the decision layer. We alleviate convoluted explanations of the conventional visualization approach by employing class information, thereby isolating mixed properties. This process offers more human-interpretable explanations for model errors without altering the trained network or requiring additional data. Furthermore, our framework mitigates false correlations learned from a dataset under a stochastic perspective, modifying decisions for the neurons considered as the main causes. We validate the effectiveness of our framework by addressing false correlations and improving inferences for classes with the worst performance in real-world settings. Moreover, we demonstrate that NeuroInspect helps debug the mistakes of DL models through evaluation for human understanding. The code is openly available at https://github.com/yeongjoonJu/NeuroInspect.

Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This risk motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse to comply with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a systematic way. XSTest comprises 250 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with, and 200 unsafe prompts as contrasts that models, for most applications, should refuse. We describe XSTest's creation and composition, and then use the test suite to highlight systematic failure modes in state-of-the-art language models as well as more general challenges in building safer language models.

In this work, we propose a novel adversarial defence mechanism for image classification - CARSO - blending the paradigms of adversarial training and adversarial purification in a mutually-beneficial, robustness-enhancing way. The method builds upon an adversarially-trained classifier, and learns to map its internal representation associated with a potentially perturbed input onto a distribution of tentative clean reconstructions. Multiple samples from such distribution are classified by the adversarially-trained model itself, and an aggregation of its outputs finally constitutes the robust prediction of interest. Experimental evaluation by a well-established benchmark of varied, strong adaptive attacks, across different image datasets and classifier architectures, shows that CARSO is able to defend itself against foreseen and unforeseen threats, including adaptive end-to-end attacks devised for stochastic defences. Paying a tolerable clean accuracy toll, our method improves by a significant margin the state of the art for CIFAR-10 and CIFAR-100 $\ell_\infty$ robust classification accuracy against AutoAttack. Code and pre-trained models are available at https://github.com/emaballarin/CARSO .

Existing question answering (QA) systems owe much of their success to large, high-quality training data. Such annotation efforts are costly, and the difficulty compounds in the cross-lingual setting. Therefore, prior cross-lingual QA work has focused on releasing evaluation datasets, and then applying zero-shot methods as baselines. This work proposes a synthetic data generation method for cross-lingual QA which leverages indirect supervision from existing parallel corpora. Our method termed PAXQA (Projecting annotations for cross-lingual (x) QA) decomposes cross-lingual QA into two stages. First, we apply a question generation (QG) model to the English side. Second, we apply annotation projection to translate both the questions and answers. To better translate questions, we propose a novel use of lexically-constrained machine translation, in which constrained entities are extracted from the parallel bitexts. We apply PAXQA to generate cross-lingual QA examples in 4 languages (662K examples total), and perform human evaluation on a subset to create validation and test splits. We then show that models fine-tuned on these datasets outperform prior synthetic data generation models over several extractive QA datasets. The largest performance gains are for directions with non-English questions and English contexts. Ablation studies show that our dataset generation method is relatively robust to noise from automatic word alignments, showing the sufficient quality of our generations. To facilitate follow-up work, we release our code and datasets at https://github.com/manestay/paxqa .

Adversarial examples are inputs to machine learning models that an attacker has intentionally designed to confuse the model into making a mistake. Such examples pose a serious threat to the applicability of machine-learning-based systems, especially in life- and safety-critical domains. To address this problem, the area of adversarial robustness investigates mechanisms behind adversarial attacks and defenses against these attacks. This survey reviews a particular subset of this literature that focuses on investigating properties of training data in the context of model robustness under evasion attacks. It first summarizes the main properties of data leading to adversarial vulnerability. It then discusses guidelines and techniques for improving adversarial robustness by enhancing the data representation and learning procedures, as well as techniques for estimating robustness guarantees given particular data. Finally, it discusses gaps of knowledge and promising future research directions in this area.

We study sequential decision making problems aimed at maximizing the expected total reward while satisfying a constraint on the expected total utility. We employ the natural policy gradient method to solve the discounted infinite-horizon optimal control problem for Constrained Markov Decision Processes (constrained MDPs). Specifically, we propose a new Natural Policy Gradient Primal-Dual (NPG-PD) method that updates the primal variable via natural policy gradient ascent and the dual variable via projected sub-gradient descent. Although the underlying maximization involves a nonconcave objective function and a nonconvex constraint set, under the softmax policy parametrization we prove that our method achieves global convergence with sublinear rates regarding both the optimality gap and the constraint violation. Such convergence is independent of the size of the state-action space, i.e., it is dimension-free. Furthermore, for log-linear and general smooth policy parametrizations, we establish sublinear convergence rates up to a function approximation error caused by restricted policy parametrization. We also provide convergence and finitesample complexity guarantees for two sample-based NPG-PD algorithms. Finally, we use computational experiments to showcase the merits and the effectiveness of our approach. Keywords: Constrained Markov decision processes; Natural policy gradient; Constrained nonconvex optimization; Method of Lagrange multipliers; Primal-dual algorithms.

Fake news is a growing challenge for social networks and media. Detection of fake news always has been a problem for many years, but after the evolution of social networks and increasing speed of news dissemination in recent years has been considered again. There are several approaches to solving this problem, one of which is to detect fake news based on its text style using deep neural networks. In recent years, one of the most used forms of deep neural networks for natural language processing is transfer learning with transformers. BERT is one of the most promising transformers who outperforms other models in many NLP benchmarks. This article, we introduce MWPBert, which uses two parallel BERT networks to perform veracity detection on full-text news articles. One of the BERT networks encodes news headline, and another encodes news body. Since the input length of the BERT network is limited and constant and the news body is usually a long text, we cannot fed the whole news text into the BERT. Therefore, using the MaxWorth algorithm, we selected the part of the news text that is more valuable for fact-checking, and fed it into the BERT network. Finally, we encode the output of the two BERT networks to an output network to classify the news. The experiment results showed that the proposed model outperformed previous models in terms of accuracy and other performance measures.

Single neurons in neural networks are often interpretable in that they represent individual, intuitively meaningful features. However, many neurons exhibit mixed selectivity, i.e., they represent multiple unrelated features. A recent hypothesis proposes that features in deep networks may be represented in superposition, i.e., on non-orthogonal axes by multiple neurons, since the number of possible interpretable features in natural data is generally larger than the number of neurons in a given network. Accordingly, we should be able to find meaningful directions in activation space that are not aligned with individual neurons. Here, we propose (1) an automated method for quantifying visual interpretability that is validated against a large database of human psychophysics judgments of neuron interpretability, and (2) an approach for finding meaningful directions in network activation space. We leverage these methods to discover directions in convolutional neural networks that are more intuitively meaningful than individual neurons, as we confirm and investigate in a series of analyses. Moreover, we apply the same method to three recent datasets of visual neural responses in the brain and find that our conclusions largely transfer to real neural data, suggesting that superposition might be deployed by the brain. This also provides a link with disentanglement and raises fundamental questions about robust, efficient and factorized representations in both artificial and biological neural systems. One of the oldest ideas in neuroscience is Cajal's single neuron doctrine (Finger, 2001) and its application to perception (Barlow, 1972), i.e., the hypothesis that individual sensory neurons encode individually meaningful features. The idea dates back to the early 1950s, when researchers began to find evidence of neurons that reliably and selectively fire in response to particular stimuli, such as dots on a contrasting background (Barlow, 1953) and lines of particular orientation and width (Hubel & Wiesel, 1959). These findings gave rise to the standard model of the ventral visual stream as a process of hierarchical feature extraction and pooling (Hubel & Wiesel, 1968; Gross et al., 1972; In this work, we adopt a pragmatic definition of feature based on human discernability, measured through psychophysics experiments (see below). For an attempt at a more formal definition see Elhage et al. (2022). Neurons in the early stages extract simple features, such as oriented lines, while neurons at later stages combine simple features to construct more complex composite features. In the highest stages, complex features are combined to yield representations of entire objects encoded by single neurons--the shape of a hand, or the face of a friend.

Radon is a carcinogenic, radioactive gas that can accumulate indoors. Indoor radon exposure at the national scale is usually estimated on the basis of extensive measurement campaigns. However, characteristics of the sample often differ from the characteristics of the population due to the large number of relevant factors such as the availability of geogenic radon or floor level. Furthermore, the sample size usually does not allow exposure estimation with high spatial resolution. We propose a model-based approach that allows a more realistic estimation of indoor radon distribution with a higher spatial resolution than a purely data-based approach. We applied a two-stage modelling approach: 1) a quantile regression forest using environmental and building data as predictors was applied to estimate the probability distribution function of indoor radon for each floor level of each residential building in Germany; (2) a probabilistic Monte Carlo sampling technique enabled the combination and population weighting of floor-level predictions. In this way, the uncertainty of the individual predictions is effectively propagated into the estimate of variability at the aggregated level. The results give an arithmetic mean of 63 Bq/m3, a geometric mean of 41 Bq/m3 and a 95 %ile of 180 Bq/m3. The exceedance probability for 100 Bq/m3 and 300 Bq/m3 are 12.5 % (10.5 million people) and 2.2 % (1.9 million people), respectively. In large cities, individual indoor radon exposure is generally lower than in rural areas, which is a due to the different distribution of the population on floor levels. The advantages of our approach are 1) an accurate exposure estimation even if the survey was not fully representative with respect to the main controlling factors, and 2) an estimate of the exposure distribution with a much higher spatial resolution than basic descriptive statistics.

Stochastic gradient descent (SGD) or stochastic approximation has been widely used in model training and stochastic optimization. While there is a huge literature on analyzing its convergence, inference on the obtained solutions from SGD has only been recently studied, yet is important due to the growing need for uncertainty quantification. We investigate two computationally cheap resampling-based methods to construct confidence intervals for SGD solutions. One uses multiple, but few, SGDs in parallel via resampling with replacement from the data, and another operates this in an online fashion. Our methods can be regarded as enhancements of established bootstrap schemes to substantially reduce the computation effort in terms of resampling requirements, while at the same time bypassing the intricate mixing conditions in existing batching methods. We achieve these via a recent so-called cheap bootstrap idea and Berry-Esseen-type bound for SGD.