Research Papers

On the Reliability and Explainability of Language Models for Program Generation

Yue Liu, Chakkrit Tantithamthavorn, Yonghui Liu +1

Recent studies have adopted pre-trained language models, such as CodeT5 and CodeGPT, for automated program generation tasks like code generation, repair, and translation. Numerous language model-based approaches have been proposed and evaluated on various benchmark datasets, demonstrating promising performance. However, there is still uncertainty about the reliability of these models, particularly their realistic ability to consistently transform code sequences. This raises the question: are these techniques sufficiently trustworthy for automated program generation? Consequently, Further research is needed to understand model logic and assess reliability and explainability. To bridge these research gaps, we conduct a thorough empirical study of eight popular language models on five representative datasets to determine the capabilities and limitations of automated program generation approaches. We further employ advanced explainable AI approaches to highlight the tokens that significantly contribute to the code transformation. We discover that state-of-the-art approaches suffer from inappropriate performance evaluation stemming from severe data duplication, causing over-optimistic results. Our explainability analysis reveals that, in various experimental scenarios, language models can recognize code grammar and structural information, but they exhibit limited robustness to changes in input sequences. Overall, more rigorous evaluation approaches and benchmarks are critical to enhance the reliability and explainability of automated program generation moving forward. Our findings provide important guidelines for this goal.

As a result of rapidly accelerating AI capabilities, over the past year, national governments and multinational bodies have announced efforts to address safety, security and ethics issues related to AI models. One high priority among these efforts is the mitigation of misuse of AI models. Many biologists have for decades sought to reduce the risks of scientific research that could lead, through accident or misuse, to high-consequence disease outbreaks. Scientists have carefully considered what types of life sciences research have the potential for both benefit and risk (dual-use), especially as scientific advances have accelerated our ability to engineer organisms and create novel variants of pathogens. Here we describe how previous experience and study by scientists and policy professionals of dual-use capabilities in the life sciences can inform risk evaluations of AI models with biological capabilities. We argue that AI model evaluations should prioritize addressing high-consequence risks (those that could cause large-scale harm to the public, such as pandemics), and that these risks should be evaluated prior to model deployment so as to allow potential biosafety and/or biosecurity measures. Scientists' experience with identifying and mitigating dual-use biological risks can help inform new approaches to evaluating biological AI models. Identifying which AI capabilities post the greatest biosecurity and biosafety concerns is necessary in order to establish targeted AI safety evaluation methods, secure these tools against accident and misuse, and avoid impeding immense potential benefits.

While shallow decision trees may be interpretable, larger ensemble models like gradient-boosted trees, which often set the state of the art in machine learning problems involving tabular data, still remain black box models. As a remedy, the Shapley value (SV) is a well-known concept in explainable artificial intelligence (XAI) research for quantifying additive feature attributions of predictions. The model-specific TreeSHAP methodology solves the exponential complexity for retrieving exact SVs from tree-based models. Expanding beyond individual feature attribution, Shapley interactions reveal the impact of intricate feature interactions of any order. In this work, we present TreeSHAP-IQ, an efficient method to compute any-order additive Shapley interactions for predictions of tree-based models. TreeSHAP-IQ is supported by a mathematical framework that exploits polynomial arithmetic to compute the interaction scores in a single recursive traversal of the tree, akin to Linear TreeSHAP. We apply TreeSHAP-IQ on state-of-the-art tree ensembles and explore interactions on well-established benchmark datasets.

WBCAtt: A White Blood Cell Dataset Annotated with Detailed Morphological Attributes

Satoshi Tsutsui, Winnie Pang, Bihan Wen

The examination of blood samples at a microscopic level plays a fundamental role in clinical diagnostics, influencing a wide range of medical conditions. For instance, an in-depth study of White Blood Cells (WBCs), a crucial component of our blood, is essential for diagnosing blood-related diseases such as leukemia and anemia. While multiple datasets containing WBC images have been proposed, they mostly focus on cell categorization, often lacking the necessary morphological details to explain such categorizations, despite the importance of explainable artificial intelligence (XAI) in medical domains. This paper seeks to address this limitation by introducing comprehensive annotations for WBC images. Through collaboration with pathologists, a thorough literature review, and manual inspection of microscopic images, we have identified 11 morphological attributes associated with the cell and its components (nucleus, cytoplasm, and granules). We then annotated ten thousand WBC images with these attributes. Moreover, we conduct experiments to predict these attributes from images, providing insights beyond basic WBC classification. As the first public dataset to offer such extensive annotations, we also illustrate specific applications that can benefit from our attribute annotations. Overall, our dataset paves the way for interpreting WBC recognition models, further advancing XAI in the fields of pathology and hematology.

In recent years, discussions of responsible AI practices have seen growing support for "participatory AI" approaches, intended to involve members of the public in the design and development of AI systems. Prior research has identified a lack of standardised methods or approaches for how to use participatory approaches in the AI development process. At present, there is a dearth of evidence on attitudes to and approaches for participation in the sites driving major AI developments: commercial AI labs. Through 12 semi-structured interviews with industry practitioners and subject-matter experts, this paper explores how commercial AI labs understand participatory AI approaches and the obstacles they have faced implementing these practices in the development of AI systems and research. We find that while interviewees view participation as a normative project that helps achieve "societally beneficial" AI systems, practitioners face numerous barriers to embedding participatory approaches in their companies: participation is expensive and resource intensive, it is "atomised" within companies, there is concern about exploitation, there is no incentive to be transparent about its adoption, and it is complicated by a lack of clear context. These barriers result in a piecemeal approach to participation that confers no decision-making power to participants and has little ongoing impact for AI labs. This papers contribution is to provide novel empirical research on the implementation of public participation in commercial AI labs, and shed light on the current challenges of using participatory approaches in this context.

Language is often considered a key aspect of human thinking, providing us with exceptional abilities to generalize, explore, plan, replan, and adapt to new situations. However, Reinforcement Learning (RL) agents are far from human-level performance in any of these abilities. We hypothesize one reason for such cognitive deficiencies is that they lack the benefits of thinking in language and that we can improve AI agents by training them to think like humans do. We introduce a novel Imitation Learning framework, Thought Cloning, where the idea is to not just clone the behaviors of human demonstrators, but also the thoughts humans have as they perform these behaviors. While we expect Thought Cloning to truly shine at scale on internet-sized datasets of humans thinking out loud while acting (e.g. online videos with transcripts), here we conduct experiments in a domain where the thinking and action data are synthetically generated. Results reveal that Thought Cloning learns much faster than Behavioral Cloning and its performance advantage grows the further out of distribution test tasks are, highlighting its ability to better handle novel situations. Thought Cloning also provides important benefits for AI Safety and Interpretability, and makes it easier to debug and improve AI. Because we can observe the agent's thoughts, we can (1) more easily diagnose why things are going wrong, making it easier to fix the problem, (2) steer the agent by correcting its thinking, or (3) prevent it from doing unsafe things it plans to do. Overall, by training agents how to think as well as behave, Thought Cloning creates safer, more powerful agents.

Human Attention-Guided Explainable Artificial Intelligence for Computer Vision Models

Guoyang Liu, Jindi Zhang, Antoni B. Chan +1

We examined whether embedding human attention knowledge into saliency-based explainable AI (XAI) methods for computer vision models could enhance their plausibility and faithfulness. We first developed new gradient-based XAI methods for object detection models to generate object-specific explanations by extending the current methods for image classification models. Interestingly, while these gradient-based methods worked well for explaining image classification models, when being used for explaining object detection models, the resulting saliency maps generally had lower faithfulness than human attention maps when performing the same task. We then developed Human Attention-Guided XAI (HAG-XAI) to learn from human attention how to best combine explanatory information from the models to enhance explanation plausibility by using trainable activation functions and smoothing kernels to maximize XAI saliency map's similarity to human attention maps. While for image classification models, HAG-XAI enhanced explanation plausibility at the expense of faithfulness, for object detection models it enhanced plausibility and faithfulness simultaneously and outperformed existing methods. The learned functions were model-specific, well generalizable to other databases.

There is a substantial and ever-growing corpus of evidence and literature exploring the impacts of Artificial intelligence (AI) technologies on society, politics, and humanity as a whole. A separate, parallel body of work has explored existential risks to humanity, including but not limited to that stemming from unaligned Artificial General Intelligence (AGI). In this paper, we problematise the notion that current and near-term artificial intelligence technologies have the potential to contribute to existential risk by acting as intermediate risk factors, and that this potential is not limited to the unaligned AGI scenario. We propose the hypothesis that certain already-documented effects of AI can act as existential risk factors, magnifying the likelihood of previously identified sources of existential risk. Moreover, future developments in the coming decade hold the potential to significantly exacerbate these risk factors, even in the absence of artificial general intelligence. Our main contribution is a (non-exhaustive) exposition of potential AI risk factors and the causal relationships between them, focusing on how AI can affect power dynamics and information security. This exposition demonstrates that there exist causal pathways from AI systems to existential risks that do not presuppose hypothetical future AI capabilities.

The increasing deployment of artificial intelligence (AI) tools to inform decision making across diverse areas including healthcare, employment, social benefits, and government policy, presents a serious risk for disabled people, who have been shown to face bias in AI implementations. While there has been significant work on analysing and mitigating algorithmic bias, the broader mechanisms of how bias emerges in AI applications are not well understood, hampering efforts to address bias where it begins. In this article, we illustrate how bias in AI-assisted decision making can arise from a range of specific design decisions, each of which may seem self-contained and non-biasing when considered separately. These design decisions include basic problem formulation, the data chosen for analysis, the use the AI technology is put to, and operational design elements in addition to the core algorithmic design. We draw on three historical models of disability common to different decision-making settings to demonstrate how differences in the definition of disability can lead to highly distinct decisions on each of these aspects of design, leading in turn to AI technologies with a variety of biases and downstream effects. We further show that the potential harms arising from inappropriate definitions of disability in fundamental design stages are further amplified by a lack of transparency and disabled participation throughout the AI design process. Our analysis provides a framework for critically examining AI technologies in decision-making contexts and guiding the development of a design praxis for disability-related AI analytics. We put forth this article to provide key questions to facilitate disability-led design and participatory development to produce more fair and equitable AI technologies in disability-related contexts.

Challenges in Applying Explainability Methods to Improve the Fairness of NLP Models

Esma Balkir, Svetlana Kiritchenko, Isar Nejadgholi +1

Motivations for methods in explainable artificial intelligence (XAI) often include detecting, quantifying and mitigating bias, and contributing to making machine learning models fairer. However, exactly how an XAI method can help in combating biases is often left unspecified. In this paper, we briefly review trends in explainability and fairness in NLP research, identify the current practices in which explainability methods are applied to detect and mitigate bias, and investigate the barriers preventing XAI methods from being used more widely in tackling fairness issues.

Human Interpretation of Saliency-based Explanation Over Text

Hendrik Schuff, Alon Jacovi, Heike Adel +2

While a lot of research in explainable AI focuses on producing effective explanations, less work is devoted to the question of how people understand and interpret the explanation. In this work, we focus on this question through a study of saliency-based explanations over textual data. Feature-attribution explanations of text models aim to communicate which parts of the input text were more influential than others towards the model decision. Many current explanation methods, such as gradient-based or Shapley value-based methods, provide measures of importance which are well-understood mathematically. But how does a person receiving the explanation (the explainee) comprehend it? And does their understanding match what the explanation attempted to communicate? We empirically investigate the effect of various factors of the input, the feature-attribution explanation, and visualization procedure, on laypeople's interpretation of the explanation. We query crowdworkers for their interpretation on tasks in English and German, and fit a GAMM model to their responses considering the factors of interest. We find that people often mis-interpret the explanations: superficial and unrelated factors, such as word length, influence the explainees' importance assignment despite the explanation communicating importance directly. We then show that some of this distortion can be attenuated: we propose a method to adjust saliencies based on model estimates of over- and under-perception, and explore bar charts as an alternative to heatmap saliency visualization. We find that both approaches can attenuate the distorting effect of specific factors, leading to better-calibrated understanding of the explanation.

There has been a surge of recent interest in sociocultural diversity in machine learning (ML) research, with researchers (i) examining the benefits of diversity as an organizational solution for alleviating problems with algorithmic bias, and (ii) proposing measures and methods for implementing diversity as a design desideratum in the construction of predictive algorithms. Currently, however, there is a gap between discussions of measures and benefits of diversity in ML, on the one hand, and the broader research on the underlying concepts of diversity and the precise mechanisms of its functional benefits, on the other. This gap is problematic because diversity is not a monolithic concept. Rather, different concepts of diversity are based on distinct rationales that should inform how we measure diversity in a given context. Similarly, the lack of specificity about the precise mechanisms underpinning diversity's potential benefits can result in uninformative generalities, invalid experimental designs, and illicit interpretations of findings. In this work, we draw on research in philosophy, psychology, and social and organizational sciences to make three contributions: First, we introduce a taxonomy of different diversity concepts from philosophy of science, and explicate the distinct epistemic and political rationales underlying these concepts. Second, we provide an overview of mechanisms by which diversity can benefit group performance. Third, we situate these taxonomies--of concepts and mechanisms--in the lifecycle of sociotechnical ML systems and make a case for their usefulness in fair and accountable ML. We do so by illustrating how they clarify the discourse around diversity in the context of ML systems, promote the formulation of more precise research questions about diversity's impact, and provide conceptual tools to further advance research and practice.