Research Papers

Ground Truth Evaluation of Neural Network Explanations with CLEVR-XAI

Leila Arras, Ahmed Osman, Wojciech Samek

The rise of deep learning in today's applications entailed an increasing need in explaining the model's decisions beyond prediction performances in order to foster trust and accountability. Recently, the field of explainable AI (XAI) has developed methods that provide such explanations for already trained neural networks. In computer vision tasks such explanations, termed heatmaps, visualize the contributions of individual pixels to the prediction. So far XAI methods along with their heatmaps were mainly validated qualitatively via human-based assessment, or evaluated through auxiliary proxy tasks such as pixel perturbation, weak object localization or randomization tests. Due to the lack of an objective and commonly accepted quality measure for heatmaps, it was debatable which XAI method performs best and whether explanations can be trusted at all. In the present work, we tackle the problem by proposing a ground truth based evaluation framework for XAI methods based on the CLEVR visual question answering task. Our framework provides a (1) selective, (2) controlled and (3) realistic testbed for the evaluation of neural network explanations. We compare ten different explanation methods, resulting in new insights about the quality and properties of XAI methods, sometimes contradicting with conclusions from previous comparative studies. The CLEVR-XAI dataset and the benchmarking code can be found at https://github.com/ahmedmagdiosman/clevr-xai.

In this paper, we argue for a paradigm shift from the current model of explainable artificial intelligence (XAI), which may be counter-productive to better human decision making. In early decision support systems, we assumed that we could give people recommendations and that they would consider them, and then follow them when required. However, research found that people often ignore recommendations because they do not trust them; or perhaps even worse, people follow them blindly, even when the recommendations are wrong. Explainable artificial intelligence mitigates this by helping people to understand how and why models give certain recommendations. However, recent research shows that people do not always engage with explainability tools enough to help improve decision making. The assumption that people will engage with recommendations and explanations has proven to be unfounded. We argue this is because we have failed to account for two things. First, recommendations (and their explanations) take control from human decision makers, limiting their agency. Second, giving recommendations and explanations does not align with the cognitive processes employed by people making decisions. This position paper proposes a new conceptual framework called Evaluative AI for explainable decision support. This is a machine-in-the-loop paradigm in which decision support tools provide evidence for and against decisions made by people, rather than provide recommendations to accept or reject. We argue that this mitigates issues of over- and under-reliance on decision support tools, and better leverages human expertise in decision making.

The application of Artificial Intelligence (AI) and Machine Learning (ML) to cybersecurity challenges has gained traction in industry and academia, partially as a result of widespread malware attacks on critical systems such as cloud infrastructures and government institutions. Intrusion Detection Systems (IDS), using some forms of AI, have received widespread adoption due to their ability to handle vast amounts of data with a high prediction accuracy. These systems are hosted in the organizational Cyber Security Operation Center (CSoC) as a defense tool to monitor and detect malicious network flow that would otherwise impact the Confidentiality, Integrity, and Availability (CIA). CSoC analysts rely on these systems to make decisions about the detected threats. However, IDSs designed using Deep Learning (DL) techniques are often treated as black box models and do not provide a justification for their predictions. This creates a barrier for CSoC analysts, as they are unable to improve their decisions based on the model's predictions. One solution to this problem is to design explainable IDS (X-IDS). This survey reviews the state-of-the-art in explainable AI (XAI) for IDS, its current challenges, and discusses how these challenges span to the design of an X-IDS. In particular, we discuss black box and white box approaches comprehensively. We also present the tradeoff between these approaches in terms of their performance and ability to produce explanations. Furthermore, we propose a generic architecture that considers human-in-the-loop which can be used as a guideline when designing an X-IDS. Research recommendations are given from three critical viewpoints: the need to define explainability for IDS, the need to create explanations tailored to various stakeholders, and the need to design metrics to evaluate explanations.

Characterising Bias in Compressed Models

Sara Hooker, Nyalleng Moorosi, Gregory Clark +2

The popularity and widespread use of pruning and quantization is driven by the severe resource constraints of deploying deep neural networks to environments with strict latency, memory and energy requirements. These techniques achieve high levels of compression with negligible impact on top-line metrics (top-1 and top-5 accuracy). However, overall accuracy hides disproportionately high errors on a small subset of examples; we call this subset Compression Identified Exemplars (CIE). We further establish that for CIE examples, compression amplifies existing algorithmic bias. Pruning disproportionately impacts performance on underrepresented features, which often coincides with considerations of fairness. Given that CIE is a relatively small subset but a great contributor of error in the model, we propose its use as a human-in-the-loop auditing tool to surface a tractable subset of the dataset for further inspection or annotation by a domain expert. We provide qualitative and quantitative support that CIE surfaces the most challenging examples in the data distribution for human-in-the-loop auditing.

Large and ever-evolving technology companies continue to invest more time and resources to incorporate responsible Artificial Intelligence (AI) into production-ready systems to increase algorithmic accountability. This paper examines and seeks to offer a framework for analyzing how organizational culture and structure impact the effectiveness of responsible AI initiatives in practice. We present the results of semi-structured qualitative interviews with practitioners working in industry, investigating common challenges, ethical tensions, and effective enablers for responsible AI initiatives. Focusing on major companies developing or utilizing AI, we have mapped what organizational structures currently support or hinder responsible AI initiatives, what aspirational future processes and structures would best enable effective initiatives, and what key elements comprise the transition from current work practices to the aspirational future.

Current literature and public discourse on "trust in AI" are often focused on the principles underlying trustworthy AI, with insufficient attention paid to how people develop trust. Given that AI systems differ in their level of trustworthiness, two open questions come to the fore: how should AI trustworthiness be responsibly communicated to ensure appropriate and equitable trust judgments by different users, and how can we protect users from deceptive attempts to earn their trust? We draw from communication theories and literature on trust in technologies to develop a conceptual model called MATCH, which describes how trustworthiness is communicated in AI systems through trustworthiness cues and how those cues are processed by people to make trust judgments. Besides AI-generated content, we highlight transparency and interaction as AI systems' affordances that present a wide range of trustworthiness cues to users. By bringing to light the variety of users' cognitive processes to make trust judgments and their potential limitations, we urge technology creators to make conscious decisions in choosing reliable trustworthiness cues for target users and, as an industry, to regulate this space and prevent malicious use. Towards these goals, we define the concepts of warranted trustworthiness cues and expensive trustworthiness cues, and propose a checklist of requirements to help technology creators identify appropriate cues to use. We present a hypothetical use case to illustrate how practitioners can use MATCH to design AI systems responsibly, and discuss future directions for research and industry efforts aimed at promoting responsible trust in AI.

Organizations are rapidly deploying artificial intelligence (AI) systems to manage their workers. However, AI has been found at times to be unfair to workers. Unfairness toward workers has been associated with decreased worker effort and increased worker turnover. To avoid such problems, AI systems must be designed to support fairness and redress instances of unfairness. Despite the attention related to AI unfairness, there has not been a theoretical and systematic approach to developing a design agenda. This paper addresses the issue in three ways. First, we introduce the organizational justice theory, three different fairness types (distributive, procedural, interactional), and the frameworks for redressing instances of unfairness (retributive justice, restorative justice). Second, we review the design literature that specifically focuses on issues of AI fairness in organizations. Third, we propose a design agenda for AI fairness in organizations that applies each of the fairness types to organizational scenarios. Then, the paper concludes with implications for future research.

Explainable Artificial Intelligence (XAI) methods are typically deployed to explain and debug black-box machine learning models. However, most proposed XAI methods are black-boxes themselves and designed for images. Thus, they rely on visual interpretability to evaluate and prove explanations. In this work, we apply XAI methods previously used in the image and text-domain on time series. We present a methodology to test and evaluate various XAI methods on time series by introducing new verification techniques to incorporate the temporal dimension. We further conduct preliminary experiments to assess the quality of selected XAI method explanations with various verification methods on a range of datasets and inspecting quality metrics on it. We demonstrate that in our initial experiments, SHAP works robust for all models, but others like DeepLIFT, LRP, and Saliency Maps work better with specific architectures.

An important aspect in developing language models that interact with humans is aligning their behavior to be useful and unharmful for their human users. This is usually achieved by tuning the model in a way that enhances desired behaviors and inhibits undesired ones, a process referred to as alignment. In this paper, we propose a theoretical approach called Behavior Expectation Bounds (BEB) which allows us to formally investigate several inherent characteristics and limitations of alignment in large language models. Importantly, we prove that within the limits of this framework, for any behavior that has a finite probability of being exhibited by the model, there exist prompts that can trigger the model into outputting this behavior, with probability that increases with the length of the prompt. This implies that any alignment process that attenuates an undesired behavior but does not remove it altogether, is not safe against adversarial prompting attacks. Furthermore, our framework hints at the mechanism by which leading alignment approaches such as reinforcement learning from human feedback make the LLM prone to being prompted into the undesired behaviors. This theoretical result is being experimentally demonstrated in large scale by the so called contemporary "chatGPT jailbreaks", where adversarial users trick the LLM into breaking its alignment guardrails by triggering it into acting as a malicious persona. Our results expose fundamental limitations in alignment of LLMs and bring to the forefront the need to devise reliable mechanisms for ensuring AI safety.

The final search query for the Systematic Literature Review (SLR) was conducted on 15th July 2022. Initially, we extracted 1707 journal and conference articles from the Scopus and Web of Science databases. Inclusion and exclusion criteria were then applied, and 58 articles were selected for the SLR. The findings show four dimensions that shape the AI explanation, which are format (explanation representation format), completeness (explanation should contain all required information, including the supplementary information), accuracy (information regarding the accuracy of the explanation), and currency (explanation should contain recent information). Moreover, along with the automatic representation of the explanation, the users can request additional information if needed. We have also found five dimensions of XAI effects: trust, transparency, understandability, usability, and fairness. In addition, we investigated current knowledge from selected articles to problematize future research agendas as research questions along with possible research paths. Consequently, a comprehensive framework of XAI and its possible effects on user behavior has been developed.

"Help Me Help the AI": Understanding How Explainability Can Support Human-AI Interaction

Sunnie S. Y. Kim, Elizabeth Anne Watkins, Olga Russakovsky +2

Despite the proliferation of explainable AI (XAI) methods, little is understood about end-users' explainability needs and behaviors around XAI explanations. To address this gap and contribute to understanding how explainability can support human-AI interaction, we conducted a mixed-methods study with 20 end-users of a real-world AI application, the Merlin bird identification app, and inquired about their XAI needs, uses, and perceptions. We found that participants desire practically useful information that can improve their collaboration with the AI, more so than technical system details. Relatedly, participants intended to use XAI explanations for various purposes beyond understanding the AI's outputs: calibrating trust, improving their task skills, changing their behavior to supply better inputs to the AI, and giving constructive feedback to developers. Finally, among existing XAI approaches, participants preferred part-based explanations that resemble human reasoning and explanations. We discuss the implications of our findings and provide recommendations for future XAI design.

Modern artificial intelligence (AI) applications require large quantities of training and test data. This need creates critical challenges not only concerning the availability of such data, but also regarding its quality. For example, incomplete, erroneous, or inappropriate training data can lead to unreliable models that produce ultimately poor decisions. Trustworthy AI applications require high-quality training and test data along many quality dimensions, such as accuracy, completeness, and consistency. We explore empirically the relationship between six data quality dimensions and the performance of 19 popular machine learning algorithms covering the tasks of classification, regression, and clustering, with the goal of explaining their performance in terms of data quality. Our experiments distinguish three scenarios based on the AI pipeline steps that were fed with polluted data: polluted training data, test data, or both. We conclude the paper with an extensive discussion of our observations.