Research Papers

Robots of the future are going to exhibit increasingly human-like and super-human intelligence in a myriad of different tasks. They are also likely going to fail and be incompliant with human preferences in increasingly subtle ways. Towards the goal of achieving autonomous robots, the robot learning community has made rapid strides in applying machine learning techniques to train robots through data and interaction. This makes the study of how best to audit these algorithms for checking their compatibility with humans, pertinent and urgent. In this paper, we draw inspiration from the AI Safety and Alignment communities and make the case that we need to urgently consider ways in which we can best audit our robot learning algorithms to check for failure modes, and ensure that when operating autonomously, they are indeed behaving in ways that the human algorithm designers intend them to. We believe that this is a challenging problem that will require efforts from the entire robot learning community, and do not attempt to provide a concrete framework for auditing. Instead, we outline high-level guidance and a possible approach towards formulating this framework which we hope will serve as a useful starting point for thinking about auditing in the context of robot learning.

The increasingly widespread application of AI models motivates increased demand for explanations from a variety of stakeholders. However, this demand is ambiguous because there are many types of 'explanation' with different evaluative criteria. In the spirit of pluralism, I chart a taxonomy of types of explanation and the associated XAI methods that can address them. When we look to expose the inner mechanisms of AI models, we develop Diagnostic-explanations. When we seek to render model output understandable, we produce Explication-explanations. When we wish to form stable generalizations of our models, we produce Expectation-explanations. Finally, when we want to justify the usage of a model, we produce Role-explanations that situate models within their social context. The motivation for such a pluralistic view stems from a consideration of causes as manipulable relationships and the different types of explanations as identifying the relevant points in AI systems we can intervene upon to affect our desired changes. This paper reduces the ambiguity in use of the word 'explanation' in the field of XAI, allowing practitioners and stakeholders a useful template for avoiding equivocation and evaluating XAI methods and putative explanations.

Algorithmic Insurance

Dimitris Bertsimas, Agni Orfanoudaki

When AI systems make errors in high-stakes domains like medical diagnosis or autonomous vehicles, a single algorithmic flaw across varying operational contexts can generate highly heterogeneous losses that challenge traditional insurance assumptions. Algorithmic insurance constitutes a novel form of financial coverage for AI-induced damages, representing an emerging market that addresses algorithm-driven liability. However, insurers currently struggle to price these risks, while AI developers lack rigorous frameworks connecting system design with financial liability exposure. We analyze the connection between operational choices of binary classification performance to tail risk exposure. Using conditional value-at-risk (CVaR) to capture extreme losses, we prove that established approaches like maximizing accuracy can significantly increase worst-case losses compared to tail risk optimization, with penalties growing quadratically as thresholds deviate from optimal. We then propose a liability insurance contract structure that mandates risk-aware classification thresholds and characterize the conditions under which it creates value for AI providers. Our analysis extends to degrading model performance and human oversight scenarios. We validate our findings through a mammography case study, demonstrating that CVaR-optimal thresholds reduce tail risk up to 13-fold compared to accuracy maximization. This risk reduction enables insurance contracts to create 14-16% gains for well-calibrated firms, while poorly calibrated firms benefit up to 65% through risk transfer, mandatory recalibration, and regulatory capital relief. Unlike traditional insurance that merely transfers risk, algorithmic insurance can function as both a financial instrument and an operational governance mechanism, simultaneously enabling efficient risk transfer while improving AI safety.

Designer-User Communication for XAI: An epistemological approach to discuss XAI design

Juliana Jansen Ferreira, Mateus Monteiro

Artificial Intelligence is becoming part of any technology we use nowadays. If the AI informs people's decisions, the explanation about AI's outcomes, results, and behavior becomes a necessary capability. However, the discussion of XAI features with various stakeholders is not a trivial task. Most of the available frameworks and methods for XAI focus on data scientists and ML developers as users. Our research is about XAI for end-users of AI systems. We argue that we need to discuss XAI early in the AI-system design process and with all stakeholders. In this work, we aimed at investigating how to operationalize the discussion about XAI scenarios and opportunities among designers and developers of AI and its end-users. We took the Signifying Message as our conceptual tool to structure and discuss XAI scenarios. We experiment with its use for the discussion of a healthcare AI-System.

Matched sample selection with GANs for mitigating attribute confounding

Chandan Singh, Guha Balakrishnan, Pietro Perona

Measuring biases of vision systems with respect to protected attributes like gender and age is critical as these systems gain widespread use in society. However, significant correlations between attributes in benchmark datasets make it difficult to separate algorithmic bias from dataset bias. To mitigate such attribute confounding during bias analysis, we propose a matching approach that selects a subset of images from the full dataset with balanced attribute distributions across protected attributes. Our matching approach first projects real images onto a generative adversarial network (GAN)'s latent space in a manner that preserves semantic attributes. It then finds image matches in this latent space across a chosen protected attribute, yielding a dataset where semantic and perceptual attributes are balanced across the protected attribute. We validate projection and matching strategies with qualitative, quantitative, and human annotation experiments. We demonstrate our work in the context of gender bias in multiple open-source facial-recognition classifiers and find that bias persists after removing key confounders via matching. Code and documentation to reproduce the results here and apply the methods to new data is available at https://github.com/csinva/matching-with-gans .

Machine Learning (ML) has recently been demonstrated to rival expert-level human accuracy in prediction and detection tasks in a variety of domains, including medicine. Despite these impressive findings, however, a key barrier to the full realization of ML's potential in medical prognoses is technology acceptance. Recent efforts to produce explainable AI (XAI) have made progress in improving the interpretability of some ML models, but these efforts suffer from limitations intrinsic to their design: they work best at identifying why a system fails, but do poorly at explaining when and why a model's prediction is correct. We posit that the acceptability of ML predictions in expert domains is limited by two key factors: the machine's horizon of prediction that extends beyond human capability, and the inability for machine predictions to incorporate human intuition into their models. We propose the use of a novel ML architecture, Neural Ordinary Differential Equations (NODEs) to enhance human understanding and encourage acceptability. Our approach prioritizes human cognitive intuition at the center of the algorithm design, and offers a distribution of predictions rather than single outputs. We explain how this approach may significantly improve human-machine collaboration in prediction tasks in expert domains such as medical prognoses. We propose a model and demonstrate, by expanding a concrete example from the literature, how our model advances the vision of future hybrid Human-AI systems.

With the increasing use of artificial intelligence (AI) services and products in recent years, issues related to their trustworthiness have emerged and AI service providers need to be prepared for various risks. In this policy recommendation, we propose a risk chain model (RCModel) that supports AI service providers in proper risk assessment and control. We hope that RCModel will contribute to the realization of trustworthy AI services.

Since COVID-19 was first identified in December 2019, various public health interventions have been implemented across the world. As different measures are implemented at different countries at different times, we conduct an assessment of the relative effectiveness of the measures implemented in 18 countries and regions using data from 22/01/2020 to 02/04/2020. We compute the top one and two measures that are most effective for the countries and regions studied during the period. Two Explainable AI techniques, SHAP and ECPI, are used in our study; such that we construct (machine learning) models for predicting the instantaneous reproduction number ($R_t$) and use the models as surrogates to the real world and inputs that the greatest influence to our models are seen as measures that are most effective. Across-the-board, city lockdown and contact tracing are the two most effective measures. For ensuring $R_t<1$, public wearing face masks is also important. Mass testing alone is not the most effective measure although when paired with other measures, it can be effective. Warm temperature helps for reducing the transmission.

Learning from Learning Machines: Optimisation, Rules, and Social Norms

Travis LaCroix, Yoshua Bengio

There is an analogy between machine learning systems and economic entities in that they are both adaptive, and their behaviour is specified in a more-or-less explicit way. It appears that the area of AI that is most analogous to the behaviour of economic entities is that of morally good decision-making, but it is an open question as to how precisely moral behaviour can be achieved in an AI system. This paper explores the analogy between these two complex systems, and we suggest that a clearer understanding of this apparent analogy may help us forward in both the socio-economic domain and the AI domain: known results in economics may help inform feasible solutions in AI safety, but also known results in AI may inform economic policy. If this claim is correct, then the recent successes of deep learning for AI suggest that more implicit specifications work better than explicit ones for solving such problems.

Explaining Image Classifiers using Statistical Fault Localization

Youcheng Sun, Hana Chockler, Xiaowei Huang +1

The black-box nature of deep neural networks (DNNs) makes it impossible to understand why a particular output is produced, creating demand for "Explainable AI". In this paper, we show that statistical fault localization (SFL) techniques from software engineering deliver high quality explanations of the outputs of DNNs, where we define an explanation as a minimal subset of features sufficient for making the same decision as for the original input. We present an algorithm and a tool called DeepCover, which synthesizes a ranking of the features of the inputs using SFL and constructs explanations for the decisions of the DNN based on this ranking. We compare explanations produced by DeepCover with those of the state-of-the-art tools GradCAM, LIME, SHAP, RISE and Extremal and show that explanations generated by DeepCover are consistently better across a broad set of experiments. On a benchmark set with known ground truth, DeepCover achieves 76.7% accuracy, which is 6% better than the second best Extremal.

Recent developments in artificial intelligence and machine learning have spurred interest in the growing field of AI safety, which studies how to prevent human-harming accidents when deploying AI systems. This paper thus explores the intersection of AI safety with evolutionary computation, to show how safety issues arise in evolutionary computation and how understanding from evolutionary computational and biological evolution can inform the broader study of AI safety.

$\textit{Embedded agents}$ are not explicitly separated from their environment, lacking clear I/O channels. Such agents can reason about and modify their internal parts, which they are incentivized to shortcut or $\textit{wirehead}$ in order to achieve the maximal reward. In this paper, we provide a taxonomy of ways by which wireheading can occur, followed by a definition of wirehead-vulnerable agents. Starting from the fully dualistic universal agent AIXI, we introduce a spectrum of partially embedded agents and identify wireheading opportunities that such agents can exploit, experimentally demonstrating the results with the GRL simulation platform AIXIjs. We contextualize wireheading in the broader class of all misalignment problems - where the goals of the agent conflict with the goals of the human designer - and conjecture that the only other possible type of misalignment is specification gaming. Motivated by this taxonomy, we define wirehead-vulnerable agents as embedded agents that choose to behave differently from fully dualistic agents lacking access to their internal parts.