Large Language Models (LLMs) are trained on vast amounts of data, most of which is automatically scraped from the internet. This data includes encyclopedic documents that harbor a vast amount of general knowledge (e.g., Wikipedia) but also potentially overlap with benchmark datasets used for evaluating LLMs. Consequently, evaluating models on test splits that might have leaked into the training set is prone to misleading conclusions. To foster sound evaluation of language models, we introduce a new test dataset named RepLiQA, suited for question-answering and topic retrieval tasks. RepLiQA is a collection of five splits of test sets, four of which have not been released to the internet or exposed to LLM APIs prior to this publication. Each sample in RepLiQA comprises (1) a reference document crafted by a human annotator and depicting an imaginary scenario (e.g., a news article) absent from the internet; (2) a question about the document’s topic; (3) a ground-truth answer derived directly from the information in the document; and (4) the paragraph extracted from the reference document containing the answer. As such, accurate answers can only be generated if a model can find relevant content within the provided document. We run a large-scale benchmark comprising several state-of-the-art LLMs to uncover differences in performance across models of various types and sizes in a context-conditional language modeling setting. Released splits of RepLiQA can be found here: https://huggingface.co/datasets/ServiceNow/repliqa.

Deployed machine learning systems require some mechanism to detect out-of-distribution (OOD) inputs. Existing research mainly focuses on one type of distribution shift: detecting samples from novel classes, absent from the training set. However, real-world systems encounter a broad variety of anomalous inputs, and the OOD literature neglects this diversity. This work categorizes five distinct types of distribution shifts and critically evaluates the performance of recent OOD detection methods on each of them. We publicly release our benchmark under the name BROAD (Benchmarking Resilience Over Anomaly Diversity). We find that while these methods excel in detecting novel classes, their performances are inconsistent across other types of distribution shifts. In other words, they can only reliably detect unexpected inputs that they have been specifically designed to expect. As a first step toward broad OOD detection, we learn a Gaussian mixture generative model for existing detection scores, enabling an ensemble detection approach that is more consistent and comprehensive for broad OOD detection, with improved performances over existing methods. We release code to build BROAD to facilitate a more comprehensive evaluation of novel OOD detectors.

Handling out-of-distribution (OOD) samples has become a major stake in the real-world deployment of machine learning systems. This work explores the use of self-supervised contrastive learning to the simultaneous detection of two types of OOD samples: unseen classes and adversarial perturbations. First, we pair self-supervised contrastive learning with the maximum mean discrepancy (MMD) two-sample test. This approach enables us to robustly test whether two independent sets of samples originate from the same distribution, and we demonstrate its effectiveness by discriminating between CIFAR-10 and CIFAR-10.1 with higher confidence than previous work. Motivated by this success, we introduce CADet (Contrastive Anomaly Detection), a novel method for OOD detection of single samples. CADet draws inspiration from MMD, but leverages the similarity between contrastive transformations of a same sample. CADet outperforms existing adversarial detection methods in identifying adversarially perturbed samples on ImageNet and achieves comparable performance to unseen label detection methods on two challenging benchmarks: ImageNet-O and iNaturalist. Significantly, CADet is fully self-supervised and requires neither labels for in-distribution samples nor access to OOD examples.

Empirical risk minimization (ERM) is sensitive to spurious correlations present in training data, which poses a significant risk when deploying systems trained under this paradigm in high-stake applications. While the existing literature focuses on maximizing group-balanced or worst-group accuracy, estimating these quantities is hindered by costly bias annotations. This study contends that current bias-unsupervised approaches to group robustness continue to rely on group information to achieve optimal performance. Firstly, these methods implicitly assume that all group combinations are represented during training. To illustrate this, we introduce a systematic generalization task on the MPI3D dataset and discover that current algorithms fail to improve the ERM baseline when combinations of observed attribute values are missing. Secondly, bias labels are still crucial for effective model selection, restricting the practicality of these methods in real-world scenarios. To address these limitations, we propose a revised methodology for training and validating debiased models in an entirely bias-unsupervised manner. We achieve this by employing pretrained self-supervised models to reliably extract bias information, which enables the integration of a logit adjustment training loss with our validation criterion. Our empirical analysis on synthetic and real-world tasks provides evidence that our approach overcomes the identified challenges and consistently enhances robust accuracy, attaining performance which is competitive with or outperforms that of state-of-the-art methods, which, conversely, rely on bias labels for validation.

The coronavirus disease 2019 (COVID-19) pandemic has drastically impacted life around the globe. As life returns to pre-pandemic routines, COVID-19 testing has become a key component, assuring that travellers and citizens are free from the disease. Conventional tests can be expensive, time-consuming (results can take up to 48h), and require laboratory testing. Rapid antigen testing, in turn, can generate results within 15-30 minutes and can be done at home, but research shows they achieve very poor sensitivity rates. In this paper, we propose an alternative test based on speech signals recorded at home with a portable device. It has been well-documented that the virus affects many of the speech production systems (e.g., lungs, larynx, and articulators). As such, we propose the use of new modulation spectral features and linear prediction analysis to characterize these changes and design a two-stage COVID-19 prediction system by fusing the proposed features. Experiments with three COVID-19 speech datasets (CSS, DiCOVA2, and Cambridge subset) show that the two-stage feature fusion system outperforms the benchmark systems of CSS and Cambridge datasets while maintaining lower complexity compared to DL-based systems. Furthermore, the two-stage system demonstrates higher generalizability to unseen conditions in a cross-dataset testing evaluation scheme. The generalizability and interpretability of our proposed system demonstrate the potential for accessible, low-cost, at-home COVID-19 testing.

As automatic speaker recognizer systems become mainstream, voice spoofing attacks are on the rise. Common attack strategies include replay, the use of text-to-speech synthesis, and voice conversion systems. While previouslyproposed end-to-end detection frameworks have shown to be effective in spotting attacks for one particular spoofing strategy, they have relied on different models, architectures, and speech representations, depending on the spoofing strategy. In practice, however, one does not have a priori information regarding the strategy an attacker might employ to fool a speaker recognizer, thus it is necessary to devise approaches which are able to detect attacks regardless of the strategy employed to generate them. In this work, we introduce an end-to-end ensemble based approach such that two models - previously shown to perform well on each considered attack strategy - are trained jointly, while a third model learns how to mix their outputs yielding a single score. Experimental results with replay and text-to-speech/voice conversion attacks show the proposed ensemble method achieving similar or superior performance when compared to systems specialized on each spoofing strategy separately.

Despite the growing interest in unsupervised learning, extracting meaningful knowledge from unlabelled audio remains an open challenge. To take a step in this direction, we recently proposed a problem-agnostic speech encoder (PASE), that combines a convolutional encoder followed by multiple neural networks, called workers, tasked to solve self-supervised problems (i.e., ones that do not require manual annotations as ground truth). PASE was shown to capture relevant speech information, including speaker voice-print and phonemes. This paper proposes PASE+, an improved version of PASE for robust speech recognition in noisy and reverberant environments. To this end, we employ an online speech distortion module, that contaminates the input signals with a variety of random disturbances. We then propose a revised encoder that better learns short- and long-term speech dynamics with an efficient combination of recurrent and convolutional networks. Finally, we refine the set of workers used in self-supervision to encourage better cooperation. Results on TIMIT, DIRHA and CHiME-5 show that PASE+ significantly outperforms both the previous version of PASE as well as common acoustic features. Interestingly, PASE+ learns transferable representations suitable for highly mismatched acoustic conditions.

In this contribution, we augment the metric learning setting by introducing a parametric pseudo-distance, trained jointly with the encoder. Several interpretations are thus drawn for the learned distance-like model’s output. We first show it approximates a likelihood ratio which can be used for hypothesis tests, and that it further induces a large divergence across the joint distributions of pairs of examples from the same and from different classes. Evaluation is performed under the verification setting consisting of determining whether sets of examples belong to the same class, even if such classes are novel and were never presented to the model during training. Empirical evaluation shows such method defines an end-to-end approach for the verification problem, able to attain better performance than simple scorers such as those based on cosine similarity and further outperforming widely used downstream classifiers. We further observe training is much simplified under the proposed approach compared to metric learning with actual distances, requiring no complex scheme to harvest pairs of examples.

Automatic speech processing applications often have to deal with the problem of aggregating local descriptors (i.e., representations of input speech data corresponding to specific portions across the time dimension) and turning them into a single fixed-dimension representation, known as global descriptor, on top of which downstream classification tasks can be performed. In this paper, we provide an empirical assessment of different time pooling strategies when used with state-of-the-art representation learning models. In particular, insights are provided as to when it is suitable to use simple statistics of local descriptors or when more sophisticated approaches are needed. Here, language identification is used as a case study and a database containing ten oriental languages under varying test conditions (short-duration test recordings, confusing languages, unseen languages) is used. Experiments are performed with classifiers trained on top of global descriptors to provide insights on open-set evaluation performance and show that appropriate selection of such pooling strategies yield embeddings able to outperform well-known benchmark systems as well as previously results based on attention only.

Generative Adversarial Networks (GANs) can successfully approximate a probability distribution and produce realistic samples. However, open questions such as sufficient convergence conditions and mode collapse still persist. In this paper, we build on existing work in the area by proposing a novel framework for training the generator against an ensemble of discriminator networks, which can be seen as a one-student/multiple-teachers setting. We formalize this problem within the full-information adversarial bandit framework, where we evaluate the capability of an algorithm to select mixtures of discriminators for providing the generator with feedback during learning. To this end, we propose a reward function which reflects the progress made by the generator and dynamically update the mixture weights allocated to each discriminator. We also draw connections between our algorithm and stochastic optimization methods and then show that existing approaches using multiple discriminators in literature can be recovered from our framework. We argue that less expressive discriminators are smoother and have a general coarse grained view of the modes map, which enforces the generator to cover a wide portion of the data distribution support. On the other hand, highly expressive discriminators ensure samples quality. Finally, experimental results show that our approach improves samples quality and diversity over existing baselines by effectively learning a curriculum. These results also support the claim that weaker discriminators have higher entropy improving modes coverage.

A robust speaker verification system is expected to provide high recognition accuracy not only in adverse environments but also in the presence of spoofing attacks, which renders voice spoofing detection as crucial to prevent automatic speaker verification systems from a security breach. In this work, we present anti-spoofing systems developed for tackling spoofing attacks introduced for the ASVspoof 2019 challenge. We employ frame-level descriptors such as discrete Fourier transform, as well as constant Q transform-based spectral and cepstral features as countermeasures. These descriptors are both used on their own with a spoofing detection classifier to detect spoofing attacks, or in tandem with deep bottleneck features, i.e. approximate posteriors parametrized by a neural network designed to discriminate between bonafide and spoof signals. Fisher vector encoding and i-vector representations are further learned from the frame-level descriptors of the signals. For modeling, we employ two classification strategies. We finally build an end-to-end anti-spoofing system by making use of modified versions of light convolution neural networks as well as well-known ResNets. Our primary system for the logical access task and a single end-to-end system for the case of physical access we attain significant improvements over two baseline systems.

This article presents a novel approach for learning domain-invariant speaker embeddings using Generative Adversarial Networks. The main idea is to confuse a domain discriminator so that it cannot tell if embeddings are from the source or target domains. We train several GAN variants using our proposed framework and apply them to the speaker verification task. On the challenging NIST-SRE 2016 dataset, we are able to match the performance of a strong baseline x-vector system. In contrast to the the baseline systems which are dependent on dimensionality reduction (LDA) and an external classifier (PLDA), our proposed speaker embeddings can be scored using simple cosine distance. This is achieved by optimizing our models end-to-end, using an angular margin loss function. Furthermore, we are able to significantly boost verification performance by averaging our different GAN models at the score level, achieving a relative improvement of 7.2% over the baseline.

Recent literature has demonstrated promising results for training Generative Adversarial Networks by employing a set of discriminators, in contrast to the traditional game involving one generator against a single adversary. Such methods perform single-objective optimization on some simple consolidation of the losses, e.g. an arithmetic average. In this work, we revisit the multiple-discriminator setting by framing the simultaneous minimization of losses provided by different models as a multi-objective optimization problem. Specifically, we evaluate the performance of multiple gradient descent and the hypervolume maximization algorithm on a number of different datasets. Moreover, we argue that the previously proposed methods and hypervolume maximization can all be seen as variations of multiple gradient descent in which the update direction can be computed efficiently. Our results indicate that hypervolume maximization presents a better compromise between sample quality and computational cost than previous methods.

Assessment of mental workload in real world conditions is key to ensure the performance of workers executing tasks which demand sustained attention. Previous literature has employed electroencephalography (EEG) to this end. However, EEG correlates of mental workload vary across subjects and physical strain, thus making it difficult to devise models capable of simultaneously presenting reliable performance across users. The field of domain adaptation (DA) aims at developing methods that allow for generalization across different domains by learning domain-invariant representations. Such DA methods, however, rely on the so-called covariate shift assumption, which typically does not hold for EEG-based applications. As such, in this paper we propose a way to measure the statistical (marginal and conditional) shift observed on data obtained from different users and use this measure to quantitatively assess the effectiveness of different adaptation strategies. In particular, we use EEG data collected from individuals performing a mental task while running in a treadmill and explore the effects of different normalization strategies commonly used to mitigate cross-subject variability. We show the effects that different normalization schemes have on statistical shifts and their relationship with the accuracy of mental workload prediction as assessed on unseen participants at train time.

Modern applications of artificial neural networks have yielded remarkable performance gains in a wide range of tasks. However, recent studies have discovered that such modelling strategy is vulnerable to Adversarial Examples, i.e. examples with subtle perturbations often too small and imperceptible to humans, but that can easily fool neural networks. Defense techniques against adversarial examples have been proposed, but ensuring robust performance against varying or novel types of attacks remains an open problem. In this work, we focus on the detection setting, in which case attackers become identifiable while models remain vulnerable. Particularly, we employ the decision layer of independently trained models as features for posterior detection. The proposed framework does not require any prior knowledge of adversarial examples generation techniques, and can be directly employed along with unmodified off-the-shelf models. Experiments on the standard MNIST and CIFAR10 datasets deliver empirical evidence that such detection approach generalizes well across not only different adversarial examples generation methods but also quality degradation attacks. Non-linear binary classifiers trained on top of our proposed features can achieve a high detection rate (>90%) in a set of white-box attacks and maintain such performance when tested against unseen attacks.