While deep neural network (DNN)-based perception models are useful for many applications, these models are black boxes and their outputs are not yet well understood. To confidently enable a real-world, decision-making system to utilize such a perception model without human intervention, we must enable the system to reason about the perception model's level of competency and respond appropriately when the model is incompetent. In order for the system to make an intelligent decision about the appropriate action when the model is incompetent, it would be useful for the system to understand why the model is incompetent. We explore five novel methods for identifying regions in the input image contributing to low model competency, which we refer to as image cropping, segment masking, pixel perturbation, competency gradients, and reconstruction loss. We assess the ability of these five methods to identify unfamiliar objects, recognize regions associated with unseen classes, and identify unexplored areas in an environment. We find that the competency gradients and reconstruction loss methods show great promise in identifying regions associated with low model competency, particularly when aspects of the image that are unfamiliar to the perception model are causing this reduction in competency. Both of these methods boast low computation times and high levels of accuracy in detecting image regions that are unfamiliar to the model, allowing them to provide potential utility in decision-making pipelines. The code for reproducing our methods and results is available on GitHub: https://github.com/sarapohland/explainable-competency.

Large Language Models (LLMs) have emerged as a new paradigm for embodied reasoning and control, most recently by generating robot policy code that utilizes a custom library of vision and control primitive skills. However, prior arts fix their skills library and steer the LLM with carefully hand-crafted prompt engineering, limiting the agent to a stationary range of addressable tasks. In this work, we introduce LRLL, an LLM-based lifelong learning agent that continuously grows the robot skill library to tackle manipulation tasks of ever-growing complexity. LRLL achieves this with four novel contributions: 1) a soft memory module that allows dynamic storage and retrieval of past experiences to serve as context, 2) a self-guided exploration policy that proposes new tasks in simulation, 3) a skill abstractor that distills recent experiences into new library skills, and 4) a lifelong learning algorithm for enabling human users to bootstrap new skills with minimal online interaction. LRLL continuously transfers knowledge from the memory to the library, building composable, general and interpretable policies, while bypassing gradient-based optimization, thus relieving the learner from catastrophic forgetting. Empirical evaluation in a simulated tabletop environment shows that LRLL outperforms end-to-end and vanilla LLM approaches in the lifelong setup while learning skills that are transferable to the real world. Project material will become available at the webpage https://gtziafas.github.io/LRLL_project.

This study proposes a unified theory and statistical learning approach for traffic conflict detection, addressing the long-existing call for a consistent and comprehensive methodology to evaluate the collision risk emerged in road user interactions. The proposed theory assumes a context-dependent probabilistic collision risk and frames conflict detection as estimating the risk by statistical learning from observed proximities and contextual variables. Three primary tasks are integrated: representing interaction context from selected observables, inferring proximity distributions in different contexts, and applying extreme value theory to relate conflict intensity with conflict probability. As a result, this methodology is adaptable to various road users and interaction scenarios, enhancing its applicability without the need for pre-labelled conflict data. Demonstration experiments are executed using real-world trajectory data, with the unified metric trained on lane-changing interactions on German highways and applied to near-crash events from the 100-Car Naturalistic Driving Study in the U.S. The experiments demonstrate the methodology's ability to provide effective collision warnings, generalise across different datasets and traffic environments, cover a broad range of conflicts, and deliver a long-tailed distribution of conflict intensity. This study contributes to traffic safety by offering a consistent and explainable methodology for conflict detection applicable across various scenarios. Its societal implications include enhanced safety evaluations of traffic infrastructures, more effective collision warning systems for autonomous and driving assistance systems, and a deeper understanding of road user behaviour in different traffic conditions, contributing to a potential reduction in accident rates and improving overall traffic safety.

The ability to properly benchmark model performance in the face of spurious correlations is important to both build better predictors and increase confidence that models are operating as intended. We demonstrate that characterizing (as opposed to simply quantifying) model mistakes across subgroups is pivotal to properly reflect model biases, which are ignored by standard metrics such as worst-group accuracy or accuracy gap. Inspired by the hypothesis testing framework, we introduce SkewSize, a principled and flexible metric that captures bias from mistakes in a model's predictions. It can be used in multi-class settings or generalised to the open vocabulary setting of generative models. SkewSize is an aggregation of the effect size of the interaction between two categorical variables: the spurious variable representing the bias attribute and the model's prediction. We demonstrate the utility of SkewSize in multiple settings including: standard vision models trained on synthetic data, vision models trained on ImageNet, and large scale vision-and-language models from the BLIP-2 family. In each case, the proposed SkewSize is able to highlight biases not captured by other metrics, while also providing insights on the impact of recently proposed techniques, such as instruction tuning.

The use of AI technologies is percolating into the secure development of software-based systems, with an increasing trend of composing AI-based subsystems (with uncertain levels of performance) into automated pipelines. This presents a fundamental research challenge and poses a serious threat to safety-critical domains (e.g., aviation). Despite the existing knowledge about uncertainty in risk analysis, no previous work has estimated the uncertainty of AI-augmented systems given the propagation of errors in the pipeline. We provide the formal underpinnings for capturing uncertainty propagation, develop a simulator to quantify uncertainty, and evaluate the simulation of propagating errors with two case studies. We discuss the generalizability of our approach and present policy implications and recommendations for aviation. Future work includes extending the approach and investigating the required metrics for validation in the aviation domain.

Published research highlights the presence of demographic bias in automated facial attribute classification. The proposed bias mitigation techniques are mostly based on supervised learning, which requires a large amount of labeled training data for generalizability and scalability. However, labeled data is limited, requires laborious annotation, poses privacy risks, and can perpetuate human bias. In contrast, self-supervised learning (SSL) capitalizes on freely available unlabeled data, rendering trained models more scalable and generalizable. However, these label-free SSL models may also introduce biases by sampling false negative pairs, especially at low-data regimes 200K images) under low compute settings. Further, SSL-based models may suffer from performance degradation due to a lack of quality assurance of the unlabeled data sourced from the web. This paper proposes a fully self-supervised pipeline for demographically fair facial attribute classifiers. Leveraging completely unlabeled data pseudolabeled via pre-trained encoders, diverse data curation techniques, and meta-learning-based weighted contrastive learning, our method significantly outperforms existing SSL approaches proposed for downstream image classification tasks. Extensive evaluations on the FairFace and CelebA datasets demonstrate the efficacy of our pipeline in obtaining fair performance over existing baselines. Thus, setting a new benchmark for SSL in the fairness of facial attribute classification.

This study proposes an effective method to predict technology development from an evolutionary perspective. Product evolution is the result of technological evolution and market selection. A phylogenetic network is the main method to study product evolution. The formation of the dominant design determines the trajectory of technology development. How to predict future dominant design has become a key issue in technology forecasting and new product development. We define the dominant product and use machine learning methods, combined with product evolutionary theory, to construct a Fully Connected Phylogenetic Network dataset to effectively predict the future dominant design.

Language models (LMs) are trained on vast amounts of text data, which may include private and copyrighted content. Data owners may request the removal of their data from a trained model due to privacy or copyright concerns. However, exactly unlearning only these datapoints (i.e., retraining with the data removed) is intractable in modern-day models. This has led to the development of many approximate unlearning algorithms. The evaluation of the efficacy of these algorithms has traditionally been narrow in scope, failing to precisely quantify the success and practicality of the algorithm from the perspectives of both the model deployers and the data owners. We address this issue by proposing MUSE, a comprehensive machine unlearning evaluation benchmark that enumerates six diverse desirable properties for unlearned models: (1) no verbatim memorization, (2) no knowledge memorization, (3) no privacy leakage, (4) utility preservation on data not intended for removal, (5) scalability with respect to the size of removal requests, and (6) sustainability over sequential unlearning requests. Using these criteria, we benchmark how effectively eight popular unlearning algorithms on 7B-parameter LMs can unlearn Harry Potter books and news articles. Our results demonstrate that most algorithms can prevent verbatim memorization and knowledge memorization to varying degrees, but only one algorithm does not lead to severe privacy leakage. Furthermore, existing algorithms fail to meet deployer's expectations because they often degrade general model utility and also cannot sustainably accommodate successive unlearning requests or large-scale content removal. Our findings identify key issues with the practicality of existing unlearning algorithms on language models, and we release our benchmark to facilitate further evaluations: muse-bench.github.io

The objective of this study was to develop an automated pipeline that enhances thyroid disease classification using thyroid scintigraphy images, aiming to decrease assessment time and increase diagnostic accuracy. Anterior thyroid scintigraphy images from 2,643 patients were collected and categorized into diffuse goiter (DG), multinodal goiter (MNG), and thyroiditis (TH) based on clinical reports, and then segmented by an expert. A ResUNet model was trained to perform auto-segmentation. Radiomic features were extracted from both physician (scenario 1) and ResUNet segmentations (scenario 2), followed by omitting highly correlated features using Spearman's correlation, and feature selection using Recursive Feature Elimination (RFE) with XGBoost as the core. All models were trained under leave-one-center-out cross-validation (LOCOCV) scheme, where nine instances of algorithms were iteratively trained and validated on data from eight centers and tested on the ninth for both scenarios separately. Segmentation performance was assessed using the Dice similarity coefficient (DSC), while classification performance was assessed using metrics, such as precision, recall, F1-score, accuracy, area under the Receiver Operating Characteristic (ROC AUC), and area under the precision-recall curve (PRC AUC). ResUNet achieved DSC values of 0.84$\pm$0.03, 0.71$\pm$0.06, and 0.86$\pm$0.02 for MNG, TH, and DG, respectively. Classification in scenario 1 achieved an accuracy of 0.76$\pm$0.04 and a ROC AUC of 0.92$\pm$0.02 while in scenario 2, classification yielded an accuracy of 0.74$\pm$0.05 and a ROC AUC of 0.90$\pm$0.02. The automated pipeline demonstrated comparable performance to physician segmentations on several classification metrics across different classes, effectively reducing assessment time while maintaining high diagnostic accuracy. Code available at: https://github.com/ahxmeds/thyroidiomics.git.

Kernel ridge regression (KRR) is a fundamental computational tool, appearing in problems that range from computational chemistry to health analytics, with a particular interest due to its starring role in Gaussian process regression. However, it is challenging to scale KRR solvers to large datasets: with $n$ training points, a direct solver (i.e., Cholesky decomposition) uses $O(n^2)$ storage and $O(n^3)$ flops. Iterative methods for KRR, such as preconditioned conjugate gradient (PCG), avoid the cubic scaling of direct solvers and often use low-rank preconditioners; a rank $r$ preconditioner uses $O(rn)$ storage and each iteration requires $O(n^2)$ flops. To reduce the storage and iteration complexity of iterative solvers for KRR, we propose ASkotch ($\textbf{A}$ccelerated $\textbf{s}$calable $\textbf{k}$ernel $\textbf{o}$p$\textbf{t}$imization using block $\textbf{c}$oordinate descent with $\textbf{H}$essian preconditioning). For a given block size $|b| << n$, each iteration of ASkotch uses $O(r|b| + n)$ storage and $O(n|b|)$ flops, so ASkotch scales better than Cholesky decomposition and PCG. We prove that ASkotch obtains linear convergence to the optimum, with the convergence rate depending on the square roots of the $\textit{preconditioned}$ block condition numbers. Furthermore, we solve KRR problems that were considered to be impossibly large while using limited computational resources: we show that ASkotch outperforms PCG methods with respect to generalization error on large-scale KRR (up to $n = 10^8$) and KRR classification tasks (up to $n = 10^7$) while running each of our experiments on $\textit{a single 12 GB Titan V GPU}$. Our work opens up the possibility of as-yet-unimagined applications of KRR across a wide range of disciplines.