Research Papers

A Framework for Automated Measurement of Responsible AI Harms in Generative AI Applications

Ahmed Magooda, Alec Helyar, Kyle Jackson +14

We present a framework for the automated measurement of responsible AI (RAI) metrics for large language models (LLMs) and associated products and services. Our framework for automatically measuring harms from LLMs builds on existing technical and sociotechnical expertise and leverages the capabilities of state-of-the-art LLMs, such as GPT-4. We use this framework to run through several case studies investigating how different LLMs may violate a range of RAI-related principles. The framework may be employed alongside domain-specific sociotechnical expertise to create measurements for new harm areas in the future. By implementing this framework, we aim to enable more advanced harm measurement efforts and further the responsible use of LLMs.

Cardiovascular diseases remain the leading global cause of mortality. Age is an important covariate whose effect is most easily investigated in a healthy cohort to properly distinguish the former from disease-related changes. Traditionally, most of such insights have been drawn from the analysis of electrocardiogram (ECG) feature changes in individuals as they age. However, these features, while informative, may potentially obscure underlying data relationships. In this paper we present the following contributions: (1) We employ a deep-learning model and a tree-based model to analyze ECG data from a robust dataset of healthy individuals across varying ages in both raw signals and ECG feature format. (2) We use explainable AI methods to identify the most discriminative ECG features across age groups.(3) Our analysis with tree-based classifiers reveals age-related declines in inferred breathing rates and identifies notably high SDANN values as indicative of elderly individuals, distinguishing them from younger adults. (4) Furthermore, the deep-learning model underscores the pivotal role of the P-wave in age predictions across all age groups, suggesting potential changes in the distribution of different P-wave types with age. These findings shed new light on age-related ECG changes, offering insights that transcend traditional feature-based approaches.

As the capabilities of large machine learning models continue to grow, and as the autonomy afforded to such models continues to expand, the spectre of a new adversary looms: the models themselves. The threat that a model might behave in a seemingly reasonable manner, while secretly and subtly modifying its behavior for ulterior reasons is often referred to as deceptive alignment in the AI Safety & Alignment communities. Consequently, we call this new direction Deceptive Alignment Monitoring. In this work, we identify emerging directions in diverse machine learning subfields that we believe will become increasingly important and intertwined in the near future for deceptive alignment monitoring, and we argue that advances in these fields present both long-term challenges and new research opportunities. We conclude by advocating for greater involvement by the adversarial machine learning community in these emerging directions.

International Institutions for Advanced AI

Lewis Ho, Joslyn Barnhart, Robert Trager +8

International institutions may have an important role to play in ensuring advanced AI systems benefit humanity. International collaborations can unlock AI's ability to further sustainable development, and coordination of regulatory efforts can reduce obstacles to innovation and the spread of benefits. Conversely, the potential dangerous capabilities of powerful and general-purpose AI systems create global externalities in their development and deployment, and international efforts to further responsible AI practices could help manage the risks they pose. This paper identifies a set of governance functions that could be performed at an international level to address these challenges, ranging from supporting access to frontier AI systems to setting international safety standards. It groups these functions into four institutional models that exhibit internal synergies and have precedents in existing organizations: 1) a Commission on Frontier AI that facilitates expert consensus on opportunities and risks from advanced AI, 2) an Advanced AI Governance Organization that sets international standards to manage global threats from advanced models, supports their implementation, and possibly monitors compliance with a future governance regime, 3) a Frontier AI Collaborative that promotes access to cutting-edge AI, and 4) an AI Safety Project that brings together leading researchers and engineers to further AI safety research. We explore the utility of these models and identify open questions about their viability.

To realize the potential benefits and mitigate potential risks of AI, it is necessary to develop a framework of governance that conforms to ethics and fundamental human values. Although several organizations have issued guidelines and ethical frameworks for trustworthy AI, without a mediating governance structure, these ethical principles will not translate into practice. In this paper, we propose a multilevel governance approach that involves three groups of interdependent stakeholders: governments, corporations, and citizens. We examine their interrelationships through dimensions of trust, such as competence, integrity, and benevolence. The levels of governance combined with the dimensions of trust in AI provide practical insights that can be used to further enhance user experiences and inform public policy related to AI.

Can Language Models Teach Weaker Agents? Teacher Explanations Improve Students via Personalization

Swarnadeep Saha, Peter Hase, Mohit Bansal

A hallmark property of explainable AI models is the ability to teach other agents, communicating knowledge of how to perform a task. While Large Language Models perform complex reasoning by generating explanations for their predictions, it is unclear whether they also make good teachers for weaker agents. To address this, we consider a student-teacher framework between two LLM agents and study if, when, and how the teacher should intervene with natural language explanations to improve the student's performance. Since communication is expensive, we define a budget such that the teacher only communicates explanations for a fraction of the data, after which the student should perform well on its own. We decompose the teaching problem along four axes: (1) if teacher's test time intervention improve student predictions, (2) when it is worth explaining a data point, (3) how the teacher should personalize explanations to better teach the student, and (4) if teacher explanations also improve students on future unexplained data. We first show that teacher LLMs can indeed intervene on student reasoning to improve their performance. Next, inspired by the Theory of Mind abilities of effective teachers, we propose building two few-shot mental models of the student. The first model defines an Intervention Function that simulates the utility of an intervention, allowing the teacher to intervene when this utility is the highest and improving student performance at lower budgets. The second model enables the teacher to personalize explanations for a particular student and outperform unpersonalized teachers. We also demonstrate that in multi-turn interactions, teacher explanations generalize and learning from explained data improves student performance on future unexplained data. Finally, we verify that misaligned teachers can lower student performance to random chance by intentionally misleading them.

Explainable Artificial Intelligence (XAI) aims to provide insights into the decision-making process of AI models, allowing users to understand their results beyond their decisions. A significant goal of XAI is to improve the performance of AI models by providing explanations for their decision-making processes. However, most XAI literature focuses on how to explain an AI system, while less attention has been given to how XAI methods can be exploited to improve an AI system. In this work, a set of well-known XAI methods typically used with Machine Learning (ML) classification tasks are investigated to verify if they can be exploited, not just to provide explanations but also to improve the performance of the model itself. To this aim, two strategies to use the explanation to improve a classification system are reported and empirically evaluated on three datasets: Fashion-MNIST, CIFAR10, and STL10. Results suggest that explanations built by Integrated Gradients highlight input features that can be effectively used to improve classification performance.

SketchXAI: A First Look at Explainability for Human Sketches

Zhiyu Qu, Yulia Gryaditskaya, Ke Li +3

This paper, for the very first time, introduces human sketches to the landscape of XAI (Explainable Artificial Intelligence). We argue that sketch as a ``human-centred'' data form, represents a natural interface to study explainability. We focus on cultivating sketch-specific explainability designs. This starts by identifying strokes as a unique building block that offers a degree of flexibility in object construction and manipulation impossible in photos. Following this, we design a simple explainability-friendly sketch encoder that accommodates the intrinsic properties of strokes: shape, location, and order. We then move on to define the first ever XAI task for sketch, that of stroke location inversion SLI. Just as we have heat maps for photos, and correlation matrices for text, SLI offers an explainability angle to sketch in terms of asking a network how well it can recover stroke locations of an unseen sketch. We offer qualitative results for readers to interpret as snapshots of the SLI process in the paper, and as GIFs on the project page. A minor but interesting note is that thanks to its sketch-specific design, our sketch encoder also yields the best sketch recognition accuracy to date while having the smallest number of parameters. The code is available at \url{https://sketchxai.github.io}.

Towards Trust of Explainable AI in Thyroid Nodule Diagnosis

Truong Thanh Hung Nguyen, Van Binh Truong, Vo Thanh Khang Nguyen +2

The ability to explain the prediction of deep learning models to end-users is an important feature to leverage the power of artificial intelligence (AI) for the medical decision-making process, which is usually considered non-transparent and challenging to comprehend. In this paper, we apply state-of-the-art eXplainable artificial intelligence (XAI) methods to explain the prediction of the black-box AI models in the thyroid nodule diagnosis application. We propose new statistic-based XAI methods, namely Kernel Density Estimation and Density map, to explain the case of no nodule detected. XAI methods' performances are considered under a qualitative and quantitative comparison as feedback to improve the data quality and the model performance. Finally, we survey to assess doctors' and patients' trust in XAI explanations of the model's decisions on thyroid nodule images.

Algorithmic fairness, the research field of making machine learning (ML) algorithms fair, is an established area in ML. As ML technologies expand their application domains, including ones with high societal impact, it becomes essential to take fairness into consideration during the building of ML systems. Yet, despite its wide range of socially sensitive applications, most work treats the issue of algorithmic bias as an intrinsic property of supervised learning, i.e., the class label is given as a precondition. Unlike prior studies in fairness, we propose an individual fairness measure and a corresponding algorithm that deal with the challenges of uncertainty arising from censorship in class labels, while enforcing similar individuals to be treated similarly from a ranking perspective, free of the Lipschitz condition in the conventional individual fairness definition. We argue that this perspective represents a more realistic model of fairness research for real-world application deployment and show how learning with such a relaxed precondition draws new insights that better explains algorithmic fairness. We conducted experiments on four real-world datasets to evaluate our proposed method compared to other fairness models, demonstrating its superiority in minimizing discrimination while maintaining predictive performance with uncertainty present.

Same Same, But Different: Conditional Multi-Task Learning for Demographic-Specific Toxicity Detection

Soumyajit Gupta, Sooyong Lee, Maria De-Arteaga +1

Algorithmic bias often arises as a result of differential subgroup validity, in which predictive relationships vary across groups. For example, in toxic language detection, comments targeting different demographic groups can vary markedly across groups. In such settings, trained models can be dominated by the relationships that best fit the majority group, leading to disparate performance. We propose framing toxicity detection as multi-task learning (MTL), allowing a model to specialize on the relationships that are relevant to each demographic group while also leveraging shared properties across groups. With toxicity detection, each task corresponds to identifying toxicity against a particular demographic group. However, traditional MTL requires labels for all tasks to be present for every data point. To address this, we propose Conditional MTL (CondMTL), wherein only training examples relevant to the given demographic group are considered by the loss function. This lets us learn group specific representations in each branch which are not cross contaminated by irrelevant labels. Results on synthetic and real data show that using CondMTL improves predictive recall over various baselines in general and for the minority demographic group in particular, while having similar overall accuracy.

Cybersickness is a common ailment associated with virtual reality (VR) user experiences. Several automated methods exist based on machine learning (ML) and deep learning (DL) to detect cybersickness. However, most of these cybersickness detection methods are perceived as computationally intensive and black-box methods. Thus, those techniques are neither trustworthy nor practical for deploying on standalone energy-constrained VR head-mounted devices (HMDs). In this work, we present an explainable artificial intelligence (XAI)-based framework, LiteVR, for cybersickness detection, explaining the model's outcome and reducing the feature dimensions and overall computational costs. First, we develop three cybersickness DL models based on long-term short-term memory (LSTM), gated recurrent unit (GRU), and multilayer perceptron (MLP). Then, we employed a post-hoc explanation, such as SHapley Additive Explanations (SHAP), to explain the results and extract the most dominant features of cybersickness. Finally, we retrain the DL models with the reduced number of features. Our results show that eye-tracking features are the most dominant for cybersickness detection. Furthermore, based on the XAI-based feature ranking and dimensionality reduction, we significantly reduce the model's size by up to 4.3x, training time by up to 5.6x, and its inference time by up to 3.8x, with higher cybersickness detection accuracy and low regression error (i.e., on Fast Motion Scale (FMS)). Our proposed lite LSTM model obtained an accuracy of 94% in classifying cybersickness and regressing (i.e., FMS 1-10) with a Root Mean Square Error (RMSE) of 0.30, which outperforms the state-of-the-art. Our proposed LiteVR framework can help researchers and practitioners analyze, detect, and deploy their DL-based cybersickness detection models in standalone VR HMDs.