While there has been a growing research interest in developing out-of-distribution (OOD) detection methods, there has been comparably little discussion around how these methods should be evaluated. Given their relevance for safe(r) AI, it is important to examine whether the basis for comparing OOD detection methods is consistent with practical needs. In this work, we take a closer look at the go-to metrics for evaluating OOD detection, and question the approach of exclusively reducing OOD detection to a binary classification task with little consideration for the detection threshold. We illustrate the limitations of current metrics (AUROC & its friends) and propose a new metric - Area Under the Threshold Curve (AUTC), which explicitly penalizes poor separation between ID and OOD samples. Scripts and data are available at https://github.com/glhr/beyond-auroc

The reliable operation of automatic systems is heavily dependent on the ability to detect faults in the underlying dynamical system. While traditional model-based methods have been widely used for fault detection, data-driven approaches have garnered increasing attention due to their ease of deployment and minimal need for expert knowledge. In this paper, we present a novel principal component analysis (PCA) method that uses occupation kernels. Occupation kernels result in feature maps that are tailored to the measured data, have inherent noise-robustness due to the use of integration, and can utilize irregularly sampled system trajectories of variable lengths for PCA. The occupation kernel PCA method is used to develop a reconstruction error approach to fault detection and its efficacy is validated using numerical simulations.

Ridge regularized sparse regression involves selecting a subset of features that explains the relationship between a design matrix and an output vector in an interpretable manner. To select the sparsity and robustness of linear regressors, techniques like leave-one-out cross-validation are commonly used for hyperparameter tuning. However, cross-validation typically increases the cost of sparse regression by several orders of magnitude. Additionally, validation metrics are noisy estimators of the test-set error, with different hyperparameter combinations giving models with different amounts of noise. Therefore, optimizing over these metrics is vulnerable to out-of-sample disappointment, especially in underdetermined settings. To address this, we make two contributions. First, we leverage the generalization theory literature to propose confidence-adjusted variants of leave-one-out that display less propensity to out-of-sample disappointment. Second, we leverage ideas from the mixed-integer literature to obtain computationally tractable relaxations of confidence-adjusted leave-one-out, thereby minimizing it without solving as many MIOs. Our relaxations give rise to an efficient coordinate descent scheme which allows us to obtain significantly lower leave-one-out errors than via other methods in the literature. We validate our theory by demonstrating we obtain significantly sparser and comparably accurate solutions than via popular methods like GLMNet and suffer from less out-of-sample disappointment. On synthetic datasets, our confidence adjustment procedure generates significantly fewer false discoveries, and improves out-of-sample performance by 2-5% compared to cross-validating without confidence adjustment. Across a suite of 13 real datasets, a calibrated version of our procedure improves the test set error by an average of 4% compared to cross-validating without confidence adjustment.

Transformers have emerged as the superior choice for face recognition tasks, but their insufficient platform acceleration hinders their application on mobile devices. In contrast, Convolutional Neural Networks (CNNs) capitalize on hardware-compatible acceleration libraries. Consequently, it has become indispensable to preserve the distillation efficacy when transferring knowledge from a Transformer-based teacher model to a CNN-based student model, known as Cross-Architecture Knowledge Distillation (CAKD). Despite its potential, the deployment of CAKD in face recognition encounters two challenges: 1) the teacher and student share disparate spatial information for each pixel, obstructing the alignment of feature space, and 2) the teacher network is not trained in the role of a teacher, lacking proficiency in handling distillation-specific knowledge. To surmount these two constraints, 1) we first introduce a Unified Receptive Fields Mapping module (URFM) that maps pixel features of the teacher and student into local features with unified receptive fields, thereby synchronizing the pixel-wise spatial information of teacher and student. Subsequently, 2) we develop an Adaptable Prompting Teacher network (APT) that integrates prompts into the teacher, enabling it to manage distillation-specific knowledge while preserving the model's discriminative capacity. Extensive experiments on popular face benchmarks and two large-scale verification sets demonstrate the superiority of our method.

Vehicle performance metrics analyze data sets consisting of subject vehicle's interactions with other road users in a nominal driving environment and provide certain performance measures as outputs. To the best of the authors' knowledge, the vehicle safety performance metrics research dates back to at least 1967. To date, there still does not exist a community-wide accepted metric or a set of metrics for vehicle safety performance assessment and justification. This issue gets further amplified with the evolving interest in Advanced Driver Assistance Systems and Automated Driving Systems. In this paper, the authors seek to perform a unified study that facilitates an improved community-wide understanding of vehicle performance metrics using the lead-vehicle interaction operational design domain as a common means of performance comparison. In particular, the authors study the diversity (including constructive formulation discrepancies and empirical performance differences) among 33 base metrics with up to 51 metric variants (with different choices of hyper-parameters) in the existing literature, published between 1967 and 2022. Two data sets are adopted for the empirical performance diversity analysis, including vehicle trajectories from normal highway driving environment and relatively high-risk incidents with collisions and near-miss cases. The analysis further implies that (i) the conceptual acceptance of a safety metric proposal can be problematic if the assumptions, conditions, and types of outcome assurance are not justified properly, and (ii) the empirical performance justification of an acceptable metric can also be problematic as a dominant consensus is not observed among metrics empirically.

Deep neural networks can detect complex data patterns and leverage them to make accurate predictions in many applications, including computer vision, natural language processing, and speech recognition, to name a few examples. These models can sometimes even outperform skilled humans [1], but they still make mistakes. Unfortunately, the severity of these mistakes is compounded by the fact that the predictions computed by neural networks are often overconfident [2], partly due to overfitting [3, 4]. Several training strategies have been developed to mitigate overfitting, including dropout [5], batch normalization [6], weight normalization [7], data augmentation [8], and early stopping [9]; the latter is the focus of this paper. Early stopping consists of continuously evaluating after each batch of stochastic gradient updates (or epoch) the predictive performance of the current model on hold-out independent data. After a large number of gradient updates, only the intermediate model achieving the best performance on the hold-out data is utilized to make predictions. This strategy is often effective at mitigating overfitting and can produce relatively accurate predictions compared to fully trained models, but it does not fully resolve overconfidence because it does not lead to models with finite-sample guarantees.

Landslides have been a regular occurrence and an alarming threat to human life and property in the era of anthropogenic global warming. An early prediction of landslide susceptibility using a data-driven approach is a demand of time. In this study, we explored the eloquent features that best describe landslide susceptibility with state-of-the-art machine learning methods. In our study, we employed state-of-the-art machine learning algorithms including XgBoost, LR, KNN, SVM, and Adaboost for landslide susceptibility prediction. To find the best hyperparameters of each individual classifier for optimized performance, we have incorporated the Grid Search method, with 10 Fold Cross-Validation. In this context, the optimized version of XgBoost outperformed all other classifiers with a Cross-validation Weighted F1 score of 94.62 %. Followed by this empirical evidence, we explored the XgBoost classifier by incorporating TreeSHAP, a game-theory-based statistical algorithm used to explain Machine Learning models, to identify eloquent features such as SLOPE, ELEVATION, TWI that complement the performance of the XGBoost classifier mostly and features such as LANDUSE, NDVI, SPI which has less effect on models performance. According to the TreeSHAP explanation of features, we selected the 9 most significant landslide causal factors out of 15. Evidently, an optimized version of XgBoost along with feature reduction by 40 % has outperformed all other classifiers in terms of popular evaluation metrics with a Cross-Validation Weighted F1 score of 95.01 % on the training and AUC score of 97 %

In recent years, a number of process-based anomaly detection schemes for Industrial Control Systems were proposed. In this work, we provide the first systematic analysis of such schemes, and introduce a taxonomy of properties that are verified by those detection systems. We then present a novel general framework to generate adversarial spoofing signals that violate physical properties of the system, and use the framework to analyze four anomaly detectors published at top security conferences. We find that three of those detectors are susceptible to a number of adversarial manipulations (e.g., spoofing with precomputed patterns), which we call Synthetic Sensor Spoofing and one is resilient against our attacks. We investigate the root of its resilience and demonstrate that it comes from the properties that we introduced. Our attacks reduce the Recall (True Positive Rate) of the attacked schemes making them not able to correctly detect anomalies. Thus, the vulnerabilities we discovered in the anomaly detectors show that (despite an original good detection performance), those detectors are not able to reliably learn physical properties of the system. Even attacks that prior work was expected to be resilient against (based on verified properties) were found to be successful. We argue that our findings demonstrate the need for both more complete attacks in datasets, and more critical analysis of process-based anomaly detectors. We plan to release our implementation as open-source, together with an extension of two public datasets with a set of Synthetic Sensor Spoofing attacks as generated by our framework.

We study joint learning of network topology and a mixed opinion dynamics, in which agents may have different update rules. Such a model captures the diversity of real individual interactions. We propose a learning algorithm based on multi-armed bandit algorithms to address the problem. The goal of the algorithm is to find each agent's update rule from several candidate rules and to learn the underlying network. At each iteration, the algorithm assumes that each agent has one of the updated rules and then modifies network estimates to reduce validation error. Numerical experiments show that the proposed algorithm improves initial estimates of the network and update rules, decreases prediction error, and performs better than other methods such as sparse linear regression and Gaussian process regression.

While a number of metrics--such as log-likelihood, userspecified This paper proposes a new metric to measure the scoring functions, and the area under the receiver calibration error of probabilistic binary classifiers, operating characteristic (ROC) curve--are used to assess the called test-based calibration error (TCE). TCE incorporates quality of probabilistic classifiers, it is usually hard or even a novel loss function based on a statistical impossible to gauge whether predictions are well-calibrated test to examine the extent to which model predictions from the values of these metrics. For assessment of calibration, differ from probabilities estimated from it is typically necessary to use a metric that measures data. It offers (i) a clear interpretation, (ii) a consistent calibration error, that is, a deviation between model predictions scale that is unaffected by class imbalance, and and probabilities of target occurrences estimated from (iii) an enhanced visual representation with repect data. The importance of assessing calibration error has been to the standard reliability diagram. In addition, we long emphasised in machine learning [Nixon et al., 2019, introduce an optimality criterion for the binning Minderer et al., 2021] and in probabilistic forecasting more procedure of calibration error metrics based on a broadly [Dawid, 1982, Degroot and Fienberg, 1983].