Input gradients have a pivotal role in a variety of applications, including adversarial attack algorithms for evaluating model robustness, explainable AI techniques for generating Saliency Maps, and counterfactual explanations. However, Saliency Maps generated by traditional neural networks are often noisy and provide limited insights. In this paper, we demonstrate that, on the contrary, the Saliency Maps of 1-Lipschitz neural networks, learnt with the dual loss of an optimal transportation problem, exhibit desirable XAI properties: They are highly concentrated on the essential parts of the image with low noise, significantly outperforming state-of-the-art explanation approaches across various models and metrics. We also prove that these maps align unprecedentedly well with human explanations on ImageNet. To explain the particularly beneficial properties of the Saliency Map for such models, we prove this gradient encodes both the direction of the transportation plan and the direction towards the nearest adversarial attack. Following the gradient down to the decision boundary is no longer considered an adversarial attack, but rather a counterfactual explanation that explicitly transports the input from one class to another. Thus, Learning with such a loss jointly optimizes the classification objective and the alignment of the gradient , i.e. the Saliency Map, to the transportation plan direction. These networks were previously known to be certifiably robust by design, and we demonstrate that they scale well for large problems and models, and are tailored for explainability using a fast and straightforward method.

Modelling and forecasting real-life human behaviour using online social media is an active endeavour of interest in politics, government, academia, and industry. Since its creation in 2006, Twitter has been proposed as a potential laboratory that could be used to gauge and predict social behaviour. During the last decade, the user base of Twitter has been growing and becoming more representative of the general population. Here we analyse this user base in the context of the 2021 Mexican Legislative Election. To do so, we use a dataset of 15 million election-related tweets in the six months preceding election day. We explore different election models that assign political preference to either the ruling parties or the opposition. We find that models using data with geographical attributes determine the results of the election with better precision and accuracy than conventional polling methods. These results demonstrate that analysis of public online data can outperform conventional polling methods, and that political analysis and general forecasting would likely benefit from incorporating such data in the immediate future. Moreover, the same Twitter dataset with geographical attributes is positively correlated with results from official census data on population and internet usage in Mexico. These findings suggest that we have reached a period in time when online activity, appropriately curated, can provide an accurate representation of offline behaviour.

We consider the problem of learning a target function corresponding to a deep, extensive-width, non-linear neural network with random Gaussian weights. We consider the asymptotic limit where the number of samples, the input dimension and the network width are proportionally large. We propose a closed-form expression for the Bayes-optimal test error, for regression and classification tasks. We further compute closed-form expressions for the test errors of ridge regression, kernel and random features regression. We find, in particular, that optimally regularized ridge regression, as well as kernel regression, achieve Bayes-optimal performances, while the logistic loss yields a near-optimal test error for classification. We further show numerically that when the number of samples grows faster than the dimension, ridge and kernel methods become suboptimal, while neural networks achieve test error close to zero from quadratically many samples.

ABSTRACT The common standard for quality evaluation of automatic speech recognition (ASR) systems is reference-based metrics such as the Word Error Rate (WER), computed using manual ground-truth transcriptions that are time-consuming and expensive to obtain. This work proposes a multi-language referenceless quality metric, which allows comparing the performance of different ASR models on a speech dataset without ground truth transcriptions. To estimate the quality of ASR hypotheses, a pre-trained language model (LM) is fine-tuned with contrastive learning in a self-supervised learning manner. In experiments conducted on several unseen test datasets consisting of outputs from top commercial ASR engines in various languages, the proposed referenceless metric obtains a much higher correlation with WER scores and their ranks than the perplexity metric from the state-of-art multi-lingual LM in all experiments, and also reduces WER Figure 1. NoRefER fine-tunes a pre-trained language model in by more than 7% when used for ensembling hypotheses.

Autonomous agents acting in real-world environments often need to reason with unknown novelties interfering with their plan execution. Novelty is an unexpected phenomenon that can alter the core characteristics, composition, and dynamics of the environment. Novelty can occur at any time in any sufficiently complex environment without any prior notice or explanation. Previous studies show that novelty has catastrophic impact on agent performance. Intelligent agents reason with an internal model of the world to understand the intricacies of their environment and to successfully execute their plans. The introduction of novelty into the environment usually renders their internal model inaccurate and the generated plans no longer applicable. Novelty is particularly prevalent in the real world where domain-specific and even predicted novelty-specific approaches are used to mitigate the novelty's impact. In this work, we demonstrate that a domain-independent AI agent designed to detect, characterize, and accommodate novelty in smaller-scope physics-based games such as Angry Birds and Cartpole can be adapted to successfully perform and reason with novelty in realistic high-fidelity simulator of the military domain.

In recent years, studies such as \cite{carmon2019unlabeled,gowal2021improving,xing2022artificial} have demonstrated that incorporating additional real or generated data with pseudo-labels can enhance adversarial training through a two-stage training approach. In this paper, we perform a theoretical analysis of the asymptotic behavior of this method in high-dimensional linear regression. While a double-descent phenomenon can be observed in ridgeless training, with an appropriate $\mathcal{L}_2$ regularization, the two-stage adversarial training achieves a better performance. Finally, we derive a shortcut cross-validation formula specifically tailored for the two-stage training method.

Localisation with Frequency-Modulated Continuous-Wave (FMCW) radar has gained increasing interest due to its inherent resistance to challenging environments. However, complex artefacts of the radar measurement process require appropriate uncertainty estimation to ensure the safe and reliable application of this promising sensor modality. In this work, we propose a multi-session map management system which constructs the best maps for further localisation based on learned variance properties in an embedding space. Using the same variance properties, we also propose a new way to introspectively reject localisation queries that are likely to be incorrect. For this, we apply robust noise-aware metric learning, which both leverages the short-timescale variability of radar data along a driven path (for data augmentation) and predicts the downstream uncertainty in metric-space-based place recognition. We prove the effectiveness of our method over extensive cross-validated tests of the Oxford Radar RobotCar and MulRan dataset. In this, we outperform the current state-of-the-art in radar place recognition and other uncertainty-aware methods when using only single nearest-neighbour queries. We also show consistent performance increases when rejecting queries based on uncertainty over a difficult test environment, which we did not observe for a competing uncertainty-aware place recognition system.

The RVL-CDIP benchmark is widely used for measuring performance on the task of document classification. Despite its widespread use, we reveal several undesirable characteristics of the RVL-CDIP benchmark. These include (1) substantial amounts of label noise, which we estimate to be 8.1% (ranging between 1.6% to 16.9% per document category); (2) presence of many ambiguous or multi-label documents; (3) a large overlap between test and train splits, which can inflate model performance metrics; and (4) presence of sensitive personally-identifiable information like US Social Security numbers (SSNs). We argue that there is a risk in using RVL-CDIP for benchmarking document classifiers, as its limited scope, presence of errors (state-of-the-art models now achieve accuracy error rates that are within our estimated label error rate), and lack of diversity make it less than ideal for benchmarking. We further advocate for the creation of a new document classification benchmark, and provide recommendations for what characteristics such a resource should include.

In this study, the main objective is to develop an algorithm capable of identifying and delineating tumor regions in breast ultrasound (BUS) and mammographic images. The technique employs two advanced deep learning architectures, namely U-Net and pretrained SAM, for tumor segmentation. The U-Net model is specifically designed for medical image segmentation and leverages its deep convolutional neural network framework to extract meaningful features from input images. On the other hand, the pretrained SAM architecture incorporates a mechanism to capture spatial dependencies and generate segmentation results. Evaluation is conducted on a diverse dataset containing annotated tumor regions in BUS and mammographic images, covering both benign and malignant tumors. This dataset enables a comprehensive assessment of the algorithm's performance across different tumor types. Results demonstrate that the U-Net model outperforms the pretrained SAM architecture in accurately identifying and segmenting tumor regions in both BUS and mammographic images. The U-Net exhibits superior performance in challenging cases involving irregular shapes, indistinct boundaries, and high tumor heterogeneity. In contrast, the pretrained SAM architecture exhibits limitations in accurately identifying tumor areas, particularly for malignant tumors and objects with weak boundaries or complex shapes. These findings highlight the importance of selecting appropriate deep learning architectures tailored for medical image segmentation. The U-Net model showcases its potential as a robust and accurate tool for tumor detection, while the pretrained SAM architecture suggests the need for further improvements to enhance segmentation performance.

Interpreting machine learning models remains a challenge, hindering their adoption in clinical settings. This paper proposes leveraging Local Interpretable Model-Agnostic Explanations (LIME) to provide interpretable descriptions of black box classification models in high-stakes sepsis detection. By analyzing misclassified instances, significant features contributing to suboptimal performance are identified. The analysis reveals regions where the classifier performs poorly, allowing the calculation of error rates within these regions. This knowledge is crucial for cautious decision-making in sepsis detection and other critical applications. The proposed approach is demonstrated using the eICU dataset, effectively identifying and visualizing regions where the classifier underperforms. By enhancing interpretability, our method promotes the adoption of machine learning models in clinical practice, empowering informed decision-making and mitigating risks in critical scenarios.