Hallucinations are a type of output error produced by deep neural networks. While this has been studied in natural language processing, they have not been researched previously in automatic speech recognition. Here, we define hallucinations in ASR as transcriptions generated by a model that are semantically unrelated to the source utterance, yet still fluent and coherent. The similarity of hallucinations to probable natural language outputs of the model creates a danger of deception and impacts the credibility of the system. We show that commonly used metrics, such as word error rates, cannot differentiate between hallucinatory and non-hallucinatory models. To address this, we propose a perturbation-based method for assessing the susceptibility of an automatic speech recognition (ASR) model to hallucination at test time, which does not require access to the training dataset. We demonstrate that this method helps to distinguish between hallucinatory and non-hallucinatory models that have similar baseline word error rates. We further explore the relationship between the types of ASR errors and the types of dataset noise to determine what types of noise are most likely to create hallucinatory outputs. We devise a framework for identifying hallucinations by analysing their semantic connection with the ground truth and their fluency. Finally, we discover how to induce hallucinations with a random noise injection to the utterance.

Backdoor attack is a common threat to deep neural networks. During testing, samples embedded with a backdoor trigger will be misclassified as an adversarial target by a backdoored model, while samples without the backdoor trigger will be correctly classified. In this paper, we present the first certified backdoor detector (CBD), which is based on a novel, adjustable conformal prediction scheme based on our proposed statistic local dominant probability. For any classifier under inspection, CBD provides 1) a detection inference, 2) the condition under which the attacks are guaranteed to be detectable for the same classification domain, and 3) a probabilistic upper bound for the false positive rate. Our theoretical results show that attacks with triggers that are more resilient to test-time noise and have smaller perturbation magnitudes are more likely to be detected with guarantees. Moreover, we conduct extensive experiments on four benchmark datasets considering various backdoor types, such as BadNet, CB, and Blend. CBD achieves comparable or even higher detection accuracy than state-of-the-art detectors, and it in addition provides detection certification. Notably, for backdoor attacks with random perturbation triggers bounded by $\ell_2\leq0.75$ which achieves more than 90\% attack success rate, CBD achieves 100\% (98\%), 100\% (84\%), 98\% (98\%), and 72\% (40\%) empirical (certified) detection true positive rates on the four benchmark datasets GTSRB, SVHN, CIFAR-10, and TinyImageNet, respectively, with low false positive rates.

Fraud detection aims to discover fraudsters deceiving other users by, for example, leaving fake reviews or making abnormal transactions. Graph-based fraud detection methods consider this task as a classification problem with two classes: frauds or normal. We address this problem using Graph Neural Networks (GNNs) by proposing a dynamic relation-attentive aggregation mechanism. Based on the observation that many real-world graphs include different types of relations, we propose to learn a node representation per relation and aggregate the node representations using a learnable attention function that assigns a different attention coefficient to each relation. Furthermore, we combine the node representations from different layers to consider both the local and global structures of a target node, which is beneficial to improving the performance of fraud detection on graphs with heterophily. By employing dynamic graph attention in all the aggregation processes, our method adaptively computes the attention coefficients for each node. Experimental results show that our method, DRAG, outperforms state-of-the-art fraud detection methods on real-world benchmark datasets.

This document outlines some of the common mistakes that occur when using machine learning, and what can be done to avoid them. Whilst it should be accessible to anyone with a basic understanding of machine learning techniques, it was originally written for research students, and focuses on issues that are of particular concern within academic research, such as the need to do rigorous comparisons and reach valid conclusions. It covers five stages of the machine learning process: what to do before model building, how to reliably build models, how to robustly evaluate models, how to compare models fairly, and how to report results.

With growing concerns regarding bias and discrimination in predictive models, the AI community has increasingly focused on assessing AI system trustworthiness. Conventionally, trustworthy AI literature relies on the probabilistic framework and calibration as prerequisites for trustworthiness. In this work, we depart from this viewpoint by proposing a novel trust framework inspired by the philosophy literature on trust. We present a precise mathematical definition of trustworthiness, termed $\mathcal{U}$-trustworthiness, specifically tailored for a subset of tasks aimed at maximizing a utility function. We argue that a model's $\mathcal{U}$-trustworthiness is contingent upon its ability to maximize Bayes utility within this task subset. Our first set of results challenges the probabilistic framework by demonstrating its potential to favor less trustworthy models and introduce the risk of misleading trustworthiness assessments. Within the context of $\mathcal{U}$-trustworthiness, we prove that properly-ranked models are inherently $\mathcal{U}$-trustworthy. Furthermore, we advocate for the adoption of the AUC metric as the preferred measure of trustworthiness. By offering both theoretical guarantees and experimental validation, AUC enables robust evaluation of trustworthiness, thereby enhancing model selection and hyperparameter tuning to yield more trustworthy outcomes.

Large scale analytics engines have become a core dependency for modern data-driven enterprises to derive business insights and drive actions. These engines support a large number of analytic jobs processing huge volumes of data on a daily basis, and workloads are often inundated with overlapping computations across multiple jobs. Reusing common computation is crucial for efficient cluster resource utilization and reducing job execution time. Detecting common computation is the first and key step for reducing this computational redundancy. However, detecting equivalence on large-scale analytics engines requires efficient and scalable solutions that are fully automated. In addition, to maximize computation reuse, equivalence needs to be detected at the semantic level instead of just the syntactic level (i.e., the ability to detect semantic equivalence of seemingly different-looking queries). Unfortunately, existing solutions fall short of satisfying these requirements. In this paper, we take a major step towards filling this gap by proposing GEqO, a portable and lightweight machine-learning-based framework for efficiently identifying semantically equivalent computations at scale. GEqO introduces two machine-learning-based filters that quickly prune out nonequivalent subexpressions and employs a semi-supervised learning feedback loop to iteratively improve its model with an intelligent sampling mechanism. Further, with its novel database-agnostic featurization method, GEqO can transfer the learning from one workload and database to another. Our extensive empirical evaluation shows that, on TPC-DS-like queries, GEqO yields significant performance gains-up to 200x faster than automated verifiers-and finds up to 2x more equivalences than optimizer and signature-based equivalence detection approaches.

This work, in a pioneering approach, attempts to build a biometric system that works purely based on the fluid mechanics governing exhaled breath. We test the hypothesis that the structure of turbulence in exhaled human breath can be exploited to build biometric algorithms. This work relies on the idea that the extrathoracic airway is unique for every individual, making the exhaled breath a biomarker. Methods including classical multi-dimensional hypothesis testing approach and machine learning models are employed in building user authentication algorithms, namely user confirmation and user identification. A user confirmation algorithm tries to verify whether a user is the person they claim to be. A user identification algorithm tries to identify a user's identity with no prior information available. A dataset of exhaled breath time series samples from 94 human subjects was used to evaluate the performance of these algorithms. The user confirmation algorithms performed exceedingly well for the given dataset with over $97\%$ true confirmation rate. The machine learning based algorithm achieved a good true confirmation rate, reiterating our understanding of why machine learning based algorithms typically outperform classical hypothesis test based algorithms. The user identification algorithm performs reasonably well with the provided dataset with over $50\%$ of the users identified as being within two possible suspects. We show surprisingly unique turbulent signatures in the exhaled breath that have not been discovered before. In addition to discussions on a novel biometric system, we make arguments to utilise this idea as a tool to gain insights into the morphometric variation of extrathoracic airway across individuals. Such tools are expected to have future potential in the area of personalised medicines.

Quantification (variously called learning to quantify or class prevalence estimation) is the area of supervised machine learning concerned with estimating the percentages of instances from a population (hereafter, a bag of examples) belonging to each of the classes of interest [González et al., 2017, Esuli et al., 2023]. Quantification finds applications in many disciplines, like the social sciences, epidemiology, or market research, in which the interest lies at the aggregate level, i.e., in which inferring characteristics of the single individual (e.g., via classification, or via regression) is of little concern since knowing group-level information is all we need. Despite the fact that binary quantification (i.e., the setting in which the classes of interest are positive vs. negative) has been, by far, the most studied scenario in the quantification literature [Card and Smith, 2018, Forman, 2008, Bella et al., 2010, Esuli and Sebastiani, 2015, Hassan et al., 2020, Moreo and Sebastiani, 2021], the truth is that many of the applications of quantification naturally arise in the multiclass regime, i.e., in cases in which there are more than two mutually exclusive classes. Examples of multiclass settings are ubiquitous, and may include the allocation of human resources to different departments in a company [Forman, 2005], the analysis of different phytoplankton species that could exist in a water sample [González et al., 2019], or the analysis of the various causes of death studied in verbal autopsies [King and Lu, 2008], to name a few. A more concrete example could consist of providing answers to questions like: "What is the percentage of tweets conveying positive, neutral, and negative opinions concerning a specific hashtag?"

Real-time safety metrics are important for the automated driving system (ADS) to assess the risk of driving situations and to assist the decision-making. Although a number of real-time safety metrics have been proposed in the literature, systematic performance evaluation of these safety metrics has been lacking. As different behavioral assumptions are adopted in different safety metrics, it is difficult to compare the safety metrics and evaluate their performance. To overcome this challenge, in this study, we propose an evaluation framework utilizing logged vehicle trajectory data, in that vehicle trajectories for both subject vehicle (SV) and background vehicles (BVs) are obtained and the prediction errors caused by behavioral assumptions can be eliminated. Specifically, we examine whether the SV is in a collision unavoidable situation at each moment, given all near-future trajectories of BVs. In this way, we level the ground for a fair comparison of different safety metrics, as a good safety metric should always alarm in advance to the collision unavoidable moment. When trajectory data from a large number of trips are available, we can systematically evaluate and compare different metrics' statistical performance. In the case study, three representative real-time safety metrics, including the time-to-collision (TTC), the PEGASUS Criticality Metric (PCM), and the Model Predictive Instantaneous Safety Metric (MPrISM), are evaluated using a large-scale simulated trajectory dataset. The proposed evaluation framework is important for researchers, practitioners, and regulators to characterize different metrics, and to select appropriate metrics for different applications. Moreover, by conducting failure analysis on moments when a safety metric failed, we can identify its potential weaknesses which are valuable for its potential refinements and improvements.

Neural networks (NNs) are increasingly used in always-on safety-critical applications deployed on hardware accelerators (NN-HAs) employing various memory technologies. Reliable continuous operation of NN is essential for safety-critical applications. During online operation, NNs are susceptible to single and multiple permanent and soft errors due to factors such as radiation, aging, and thermal effects. Explicit NN-HA testing methods cannot detect transient faults during inference, are unsuitable for always-on applications, and require extensive test vector generation and storage. Therefore, in this paper, we propose the \emph{uncertainty fingerprint} approach representing the online fault status of NN. Furthermore, we propose a dual head NN topology specifically designed to produce uncertainty fingerprints and the primary prediction of the NN in \emph{a single shot}. During the online operation, by matching the uncertainty fingerprint, we can concurrently self-test NNs with up to $100\%$ coverage with a low false positive rate while maintaining a similar performance of the primary task. Compared to existing works, memory overhead is reduced by up to $243.7$ MB, multiply and accumulate (MAC) operation is reduced by up to $10000\times$, and false-positive rates are reduced by up to $89\%$.