Research Papers

Evaluating the Faithfulness of Saliency-based Explanations for Deep Learning Models for Temporal Colour Constancy

Matteo Rizzo, Cristina Conati, Daesik Jang +1

The opacity of deep learning models constrains their debugging and improvement. Augmenting deep models with saliency-based strategies, such as attention, has been claimed to help get a better understanding of the decision-making process of black-box models. However, some recent works challenged saliency's faithfulness in the field of Natural Language Processing (NLP), questioning attention weights' adherence to the true decision-making process of the model. We add to this discussion by evaluating the faithfulness of in-model saliency applied to a video processing task for the first time, namely, temporal colour constancy. We perform the evaluation by adapting to our target task two tests for faithfulness from recent NLP literature, whose methodology we refine as part of our contributions. We show that attention fails to achieve faithfulness, while confidence, a particular type of in-model visual saliency, succeeds.

Human lives are increasingly being affected by the outcomes of automated decision-making systems and it is essential for the latter to be, not only accurate, but also fair. The literature of algorithmic fairness has grown considerably over the last decade, where most of the approaches are evaluated under the strong assumption that the train and test samples are independently and identically drawn from the same underlying distribution. However, in practice, dissimilarity between the training and deployment environments exists, which compromises the performance of the decision-making algorithm as well as its fairness guarantees in the deployment data. There is an emergent research line that studies how to preserve fairness guarantees when the data generating processes differ between the source (train) and target (test) domains, which is growing remarkably. With this survey, we aim to provide a wide and unifying overview on the topic. For such purpose, we propose a taxonomy of the existing approaches for fair classification under distribution shift, highlight benchmarking alternatives, point out the relation with other similar research fields and eventually, identify future venues of research.

This paper presented an XAI system specification and an interactive dialogue model to facilitate the creation of Explanation Experiences (EE). Such specifications combine the knowledge of XAI, domain and system experts of a use case to formalise target user groups and their explanation needs and to implement explanation strategies to address those needs. Formalising the XAI system promotes the reuse of existing explainers and known explanation needs that can be refined and evolved over time using user evaluation feedback. The abstract EE dialogue model formalised the interactions between a user and an XAI system. The resulting EE conversational chatbot is personalised to an XAI system at run-time using the knowledge captured in its XAI system specification. This seamless integration is enabled by using Behaviour Trees (BT) to conceptualise both the EE dialogue model and the explanation strategies. In the evaluation, we discussed several desirable properties of using BTs over traditionally used STMs or FSMs. BTs promote the reusability of dialogue components through the hierarchical nature of the design. Sub-trees are modular, i.e. a sub-tree is responsible for a specific behaviour, which can be designed in different levels of granularity to improve human interpretability. The EE dialogue model consists of abstract behaviours needed to capture EE, accordingly, it can be implemented as a conversational, graphical or text-based interface which caters to different domains and users. There is a significant computational cost when using BTs for modelling dialogue, which we mitigate by using memory. Overall, we find that the ability to create robust conversational pathways dynamically makes BTs a good candidate for designing and implementing conversation for creating explanation experiences.

Although prediction models for delirium, a commonly occurring condition during general hospitalization or post-surgery, have not gained huge popularity, their algorithmic bias evaluation is crucial due to the existing association between social determinants of health and delirium risk. In this context, using MIMIC-III and another academic hospital dataset, we present some initial experimental evidence showing how sociodemographic features such as sex and race can impact the model performance across subgroups. With this work, our intent is to initiate a discussion about the intersectionality effects of old age, race and socioeconomic factors on the early-stage detection and prevention of delirium using ML.

Explainable Artificial Intelligence (XAI) has recently gained a swell of interest, as many Artificial Intelligence (AI) practitioners and developers are compelled to rationalize how such AI-based systems work. Decades back, most XAI systems were developed as knowledge-based or expert systems. These systems assumed reasoning for the technical description of an explanation, with little regard for the user's cognitive capabilities. The emphasis of XAI research appears to have turned to a more pragmatic explanation approach for better understanding. An extensive area where cognitive science research may substantially influence XAI advancements is evaluating user knowledge and feedback, which are essential for XAI system evaluation. To this end, we propose a framework to experiment with generating and evaluating the explanations on the grounds of different cognitive levels of understanding. In this regard, we adopt Bloom's taxonomy, a widely accepted model for assessing the user's cognitive capability. We utilize the counterfactual explanations as an explanation-providing medium encompassed with user feedback to validate the levels of understanding about the explanation at each cognitive level and improvise the explanation generation methods accordingly.

The 4th Industrial Revolution is the culmination of the digital age. Nowadays, technologies such as robotics, nanotechnology, genetics, and artificial intelligence promise to transform our world and the way we live. Artificial Intelligence Ethics and Safety is an emerging research field that has been gaining popularity in recent years. Several private, public and non-governmental organizations have published guidelines proposing ethical principles for regulating the use and development of autonomous intelligent systems. Meta-analyses of the AI Ethics research field point to convergence on certain principles that supposedly govern the AI industry. However, little is known about the effectiveness of this form of Ethics. In this paper, we would like to conduct a critical analysis of the current state of AI Ethics and suggest that this form of governance based on principled ethical guidelines is not sufficient to norm the AI industry and its developers. We believe that drastic changes are necessary, both in the training processes of professionals in the fields related to the development of software and intelligent systems and in the increased regulation of these professionals and their industry. To this end, we suggest that law should benefit from recent contributions from bioethics, to make the contributions of AI ethics to governance explicit in legal terms.

The Influence of Explainable Artificial Intelligence: Nudging Behaviour or Boosting Capability?

Matija Franklin

This article aims to provide a theoretical account and corresponding paradigm for analysing how explainable artificial intelligence (XAI) influences people's behaviour and cognition. It uses insights from research on behaviour change. Two notable frameworks for thinking about behaviour change techniques are nudges - aimed at influencing behaviour - and boosts - aimed at fostering capability. It proposes that local and concept-based explanations are more adjacent to nudges, while global and counterfactual explanations are more adjacent to boosts. It outlines a method for measuring XAI influence and argues for the benefits of understanding it for optimal, safe and ethical human-AI collaboration.

Outlier detection tasks have been playing a critical role in AI safety. There has been a great challenge to deal with this task. Observations show that deep neural network classifiers usually tend to incorrectly classify out-of-distribution (OOD) inputs into in-distribution classes with high confidence. Existing works attempt to solve the problem by explicitly imposing uncertainty on classifiers when OOD inputs are exposed to the classifier during training. In this paper, we propose an alternative probabilistic paradigm that is both practically useful and theoretically viable for the OOD detection tasks. Particularly, we impose statistical independence between inlier and outlier data during training, in order to ensure that inlier data reveals little information about OOD data to the deep estimator during training. Specifically, we estimate the statistical dependence between inlier and outlier data through the Hilbert-Schmidt Independence Criterion (HSIC), and we penalize such metric during training. We also associate our approach with a novel statistical test during the inference time coupled with our principled motivation. Empirical results show that our method is effective and robust for OOD detection on various benchmarks. In comparison to SOTA models, our approach achieves significant improvement regarding FPR95, AUROC, and AUPR metrics. Code is available: \url{https://github.com/jylins/hood}.

In the last few years, many explanation methods based on the perturbations of input data have been introduced to improve our understanding of decisions made by black-box models. The goal of this work is to introduce a novel perturbation scheme so that more faithful and robust explanations can be obtained. Our study focuses on the impact of perturbing directions on the data topology. We show that perturbing along the orthogonal directions of the input manifold better preserves the data topology, both in the worst-case analysis of the discrete Gromov-Hausdorff distance and in the average-case analysis via persistent homology. From those results, we introduce EMaP algorithm, realizing the orthogonal perturbation scheme. Our experiments show that EMaP not only improves the explainers' performance but also helps them overcome a recently-developed attack against perturbation-based methods.

On Model Reconciliation: How to Reconcile When Robot Does not Know Human's Model?

Ho Tuan Dung, Tran Cao Son

The Model Reconciliation Problem (MRP) was introduced to address issues in explainable AI planning. A solution to a MRP is an explanation for the differences between the models of the human and the planning agent (robot). Most approaches to solving MRPs assume that the robot, who needs to provide explanations, knows the human model. This assumption is not always realistic in several situations (e.g., the human might decide to update her model and the robot is unaware of the updates). In this paper, we propose a dialog-based approach for computing explanations of MRPs under the assumptions that (i) the robot does not know the human model; (ii) the human and the robot share the set of predicates of the planning domain and their exchanges are about action descriptions and fluents' values; (iii) communication between the parties is perfect; and (iv) the parties are truthful. A solution of a MRP is computed through a dialog, defined as a sequence of rounds of exchanges, between the robot and the human. In each round, the robot sends a potential explanation, called proposal, to the human who replies with her evaluation of the proposal, called response. We develop algorithms for computing proposals by the robot and responses by the human and implement these algorithms in a system that combines imperative means with answer set programming using the multi-shot feature of clingo.

Convolutional neural networks (CNN) are known for their excellent feature extraction capabilities to enable the learning of models from data, yet are used as black boxes. An interpretation of the convolutional filtres and associated features can help to establish an understanding of CNN to distinguish various classes. In this work, we focus on the explainability of a CNN model called as cnnexplain that is used for Covid-19 and non-Covid-19 classification with a focus on the interpretability of features by the convolutional filters, and how these features contribute to classification. Specifically, we have used various explainable artificial intelligence (XAI) methods, such as visualizations, SmoothGrad, Grad-CAM, and LIME to provide interpretation of convolutional filtres, and relevant features, and their role in classification. We have analyzed the explanation of these methods for Covid-19 detection using dry cough spectrograms. Explanation results obtained from the LIME, SmoothGrad, and Grad-CAM highlight important features of different spectrograms and their relevance to classification.

More Data Can Lead Us Astray: Active Data Acquisition in the Presence of Label Bias

Yunyi Li, Maria De-Arteaga, Maytal Saar-Tsechansky

An increased awareness concerning risks of algorithmic bias has driven a surge of efforts around bias mitigation strategies. A vast majority of the proposed approaches fall under one of two categories: (1) imposing algorithmic fairness constraints on predictive models, and (2) collecting additional training samples. Most recently and at the intersection of these two categories, methods that propose active learning under fairness constraints have been developed. However, proposed bias mitigation strategies typically overlook the bias presented in the observed labels. In this work, we study fairness considerations of active data collection strategies in the presence of label bias. We first present an overview of different types of label bias in the context of supervised learning systems. We then empirically show that, when overlooking label bias, collecting more data can aggravate bias, and imposing fairness constraints that rely on the observed labels in the data collection process may not address the problem. Our results illustrate the unintended consequences of deploying a model that attempts to mitigate a single type of bias while neglecting others, emphasizing the importance of explicitly differentiating between the types of bias that fairness-aware algorithms aim to address, and highlighting the risks of neglecting label bias during data collection.