Research Papers

Can I trust my anomaly detection system? A case study based on explainable AI

Muhammad Rashid, Elvio Amparore, Enrico Ferrari +1

Generative models based on variational autoencoders are a popular technique for detecting anomalies in images in a semi-supervised context. A common approach employs the anomaly score to detect the presence of anomalies, and it is known to reach high level of accuracy on benchmark datasets. However, since anomaly scores are computed from reconstruction disparities, they often obscure the detection of various spurious features, raising concerns regarding their actual efficacy. This case study explores the robustness of an anomaly detection system based on variational autoencoder generative models through the use of eXplainable AI methods. The goal is to get a different perspective on the real performances of anomaly detectors that use reconstruction differences. In our case study we discovered that, in many cases, samples are detected as anomalous for the wrong or misleading factors.

Fairness Through Controlled (Un)Awareness in Node Embeddings

Dennis Vetter, Jasper Forth, Gemma Roig +1

Graph representation learning is central for the application of machine learning (ML) models to complex graphs, such as social networks. Ensuring `fair' representations is essential, due to the societal implications and the use of sensitive personal data. In this paper, we demonstrate how the parametrization of the \emph{CrossWalk} algorithm influences the ability to infer a sensitive attributes from node embeddings. By fine-tuning hyperparameters, we show that it is possible to either significantly enhance or obscure the detectability of these attributes. This functionality offers a valuable tool for improving the fairness of ML systems utilizing graph embeddings, making them adaptable to different fairness paradigms.

Human trust is a prerequisite to trustworthy AI adoption, yet trust remains poorly understood. Trust is often described as an attitude, but attitudes cannot be reliably measured or managed. Additionally, humans frequently conflate trust in an AI system, its machine learning (ML) technology, and its other component parts. Without fully understanding the 'leap of faith' involved in trusting ML, users cannot develop intrinsic trust in these systems. A common approach to building trust is to explain a ML model's reasoning process. However, such explanations often fail to resonate with non-experts due to the inherent complexity of ML systems and explanations are disconnected from users' own (unarticulated) mental models. This work puts forward an innovative way of directly building intrinsic trust in ML, by discerning and measuring the Leap of Faith (LoF) taken when a user decides to rely on ML. The LoF matrix captures the alignment between an ML model and a human expert's mental model. This match is rigorously and practically identified by feeding the user's data and objective function into both an ML agent and an expert-validated rules-based agent: a verified point of reference that can be tested a priori against a user's own mental model. This represents a new class of neuro-symbolic architecture. The LoF matrix reveals to the user the distance that constitutes the leap of faith between the rules-based and ML agents. For the first time, we propose trust metrics that evaluate whether users demonstrate trust through their actions rather than self-reported intent and whether such trust is deserved based on outcomes. The significance of the contribution is that it enables empirical assessment and management of ML trust drivers, to support trustworthy ML adoption. The approach is illustrated through a long-term high-stakes field study: a 3-month pilot of a multi-agent sleep-improvement system.

Visual explanations based on user-uploaded images are an effective and self-contained approach to provide transparency to Recommender Systems (RS), but intrinsic limitations of data used in this explainability paradigm cause existing approaches to use bad quality training data that is highly sparse and suffers from labelling noise. Popular training enrichment approaches like model enlargement or massive data gathering are expensive and environmentally unsustainable, thus we seek to provide better visual explanations to RS aligning with the principles of Responsible AI. In this work, we research the intersection of effective and sustainable training enrichment strategies for visual-based RS explainability models by developing three novel strategies that focus on training Data Quality: 1) selection of reliable negative training examples using Positive-unlabelled Learning, 2) transform-based data augmentation, and 3) text-to-image generative-based data augmentation. The integration of these strategies in three state-of-the-art explainability models increases 5% the performance in relevant ranking metrics of these visual-based RS explainability models without penalizing their practical long-term sustainability, as tested in multiple real-world restaurant recommendation explanation datasets.

Short-Long Policy Evaluation with Novel Actions

Hyunji Alex Nam, Yash Chandak, Emma Brunskill

From incorporating LLMs in education, to identifying new drugs and improving ways to charge batteries, innovators constantly try new strategies in search of better long-term outcomes for students, patients and consumers. One major bottleneck in this innovation cycle is the amount of time it takes to observe the downstream effects of a decision policy that incorporates new interventions. The key question is whether we can quickly evaluate long-term outcomes of a new decision policy without making long-term observations. Organizations often have access to prior data about past decision policies and their outcomes, evaluated over the full horizon of interest. Motivated by this, we introduce a new setting for short-long policy evaluation for sequential decision making tasks. Our proposed methods significantly outperform prior results on simulators of HIV treatment, kidney dialysis and battery charging. We also demonstrate that our methods can be useful for applications in AI safety by quickly identifying when a new decision policy is likely to have substantially lower performance than past policies.

Black box deep learning models trained on genomic sequences excel at predicting the outcomes of different gene regulatory mechanisms. Therefore, interpreting these models may provide novel insights into the underlying biology, supporting downstream biomedical applications. Due to their complexity, interpretable surrogate models can only be built for local explanations (e.g., a single instance). However, accomplishing this requires generating a dataset in the neighborhood of the input, which must maintain syntactic similarity to the original data while introducing semantic variability in the model's predictions. This task is challenging due to the complex sequence-to-function relationship of DNA. We propose using Genetic Programming to generate datasets by evolving perturbations in sequences that contribute to their semantic diversity. Our custom, domain-guided individual representation effectively constrains syntactic similarity, and we provide two alternative fitness functions that promote diversity with no computational effort. Applied to the RNA splicing domain, our approach quickly achieves good diversity and significantly outperforms a random baseline in exploring the search space, as shown by our proof-of-concept, short RNA sequence. Furthermore, we assess its generalizability and demonstrate scalability to larger sequences, resulting in a ~30% improvement over the baseline.

Heatmaps generated on inputs of image classification networks via explainable AI methods like Grad-CAM and LRP have been observed to resemble segmentations of input images in many cases. Consequently, heatmaps have also been leveraged for achieving weakly supervised segmentation with image-level supervision. On the other hand, losses can be imposed on differentiable heatmaps, which has been shown to serve for (1)~improving heatmaps to be more human-interpretable, (2)~regularization of networks towards better generalization, (3)~training diverse ensembles of networks, and (4)~for explicitly ignoring confounding input features. Due to the latter use case, the paradigm of imposing losses on heatmaps is often referred to as "Right for the right reasons". We unify these two lines of research by investigating semi-supervised segmentation as a novel use case for the Right for the Right Reasons paradigm. First, we show formal parallels between differentiable heatmap architectures and standard encoder-decoder architectures for image segmentation. Second, we show that such differentiable heatmap architectures yield competitive results when trained with standard segmentation losses. Third, we show that such architectures allow for training with weak supervision in the form of image-level labels and small numbers of pixel-level labels, outperforming comparable encoder-decoder models. Code is available at \url{https://github.com/Kainmueller-Lab/TW-autoencoder}.

We propose a prototype-based approach for improving explainability of softmax classifiers that provides an understandable prediction confidence, generated through stochastic sampling of prototypes, and demonstrates potential for out of distribution detection (OOD). By modifying the model architecture and training to make predictions using similarities to any set of class examples from the training dataset, we acquire the ability to sample for prototypical examples that contributed to the prediction, which provide an instance-based explanation for the model's decision. Furthermore, by learning relationships between images from the training dataset through relative distances within the model's latent space, we obtain a metric for uncertainty that is better able to detect out of distribution data than softmax confidence.

This paper proposes a framework combining Neural Ordinary Differential Equations (Neural ODEs) and robust control theory to enhance the interpretability and control of large language models (LLMs). By utilizing Neural ODEs to model the dynamic evolution of input-output relationships and introducing control mechanisms to optimize output quality, we demonstrate the effectiveness of this approach across multiple question-answer datasets. Experimental results show that the integration of Neural ODEs and control theory significantly improves output consistency and model interpretability, advancing the development of explainable AI technologies.

This second international workshop on explainable AI for the Arts (XAIxArts) brought together a community of researchers in HCI, Interaction Design, AI, explainable AI (XAI), and digital arts to explore the role of XAI for the Arts. Workshop held at the 16th ACM Conference on Creativity and Cognition (C&C 2024), Chicago, USA.

The Benefits and Risks of Transductive Approaches for AI Fairness

Muhammed Razzak, Andreas Kirsch, Yarin Gal

Recently, transductive learning methods, which leverage holdout sets during training, have gained popularity for their potential to improve speed, accuracy, and fairness in machine learning models. Despite this, the composition of the holdout set itself, particularly the balance of sensitive sub-groups, has been largely overlooked. Our experiments on CIFAR and CelebA datasets show that compositional changes in the holdout set can substantially influence fairness metrics. Imbalanced holdout sets exacerbate existing disparities, while balanced holdouts can mitigate issues introduced by imbalanced training data. These findings underline the necessity of constructing holdout sets that are both diverse and representative.

This paper reviews our previous trials of Nudge-XAI, an approach that introduces automatic biases into explanations from explainable AIs (XAIs) with the aim of leading users to better decisions, and it discusses the benefits and challenges. Nudge-XAI uses a user model that predicts the influence of providing an explanation or emphasizing it and attempts to guide users toward AI-suggested decisions without coercion. The nudge design is expected to enhance the autonomy of users, reduce the risk associated with an AI making decisions without users' full agreement, and enable users to avoid AI failures. To discuss the potential of Nudge-XAI, this paper reports a post-hoc investigation of previous experimental results using cluster analysis. The results demonstrate the diversity of user behavior in response to Nudge-XAI, which supports our aim of enhancing user autonomy. However, it also highlights the challenge of users who distrust AI and falsely make decisions contrary to AI suggestions, suggesting the need for personalized adjustment of the strength of nudges to make this approach work more generally.