As data-driven AI models achieve unprecedented feats across previously unthinkable tasks, the diminishing levels of interpretability of their increasingly complex architectures can often be sidelined in place of performance. If we are to comprehend and trust these AI models as they advance, it is clear that symbolic methods, given their unparalleled strengths in knowledge representation and reasoning, can play an important role in explaining AI models. In this paper, I discuss some of the ways in which one branch of such methods, computational argumentation, given its human-like nature, can be used to tackle this problem. I first outline a general paradigm for this area of explainable AI, before detailing a prominent methodology therein which we have pioneered. I then illustrate how this approach has been put into practice with diverse AI models and types of explanations, before looking ahead to challenges, future work and the outlook in this field.
论文检索
输入标题、作者或关键词，从 7,537 篇学术成果中精准定位
Human-Robot Alignment through Interactivity and Interpretability: Don't Assume a ``Spherical Human''
PDF ↗Interactive and interpretable robot learning can help to democratize robots, placing the power of assistive robotic systems in the hands of end-users. While machine learning-based approaches to robotics have achieved impressive results, robot learning is still a feat of costly engineering performed in controlled settings and relying upon impractical assumptions about humans. To achieve a vision in which robots can be integrated sustainably into our daily lives for robotic assistance, researchers must take a human-centered approach and develop novel approaches for human-robot alignment of robot values and behaviors. This paper amalgamates recent human factors insights and computational techniques that can support human-robot alignment through interactive and interpretable robot learning and teaming.
Models for text-to-image synthesis, such as DALL-E 2 and Stable Diffusion, have recently drawn a lot of interest from academia and the general public. These models are capable of producing high-quality images that depict a variety of concepts and styles when conditioned on textual descriptions. However, these models adopt cultural characteristics associated with specific Unicode scripts from their vast amount of training data, which may not be immediately apparent. We show that by simply inserting single non-Latin characters in the textual description, common models reflect cultural biases in their generated images. We analyze this behavior both qualitatively and quantitatively and identify a model’s text encoder as the root cause of the phenomenon. Such behavior can be interpreted as a model feature, offering users a simple way to customize the image generation and reflect their own cultural background. Yet, malicious users or service providers may also try to intentionally bias the image generation. One goal might be to create racist stereotypes by replacing Latin characters with similarly-looking characters from non-Latin scripts, so-called homoglyphs. To mitigate such unnoticed script attacks, we propose a novel homoglyph unlearning method to fine-tune a text encoder, making it robust against homoglyph manipulations.
Counterfactual explanations (CEs) are advocated as being ideally suited to providing algorithmic recourse for subjects affected by the predictions of machine learning models. While CEs can be beneficial to affected individuals, recent work has exposed severe issues related to the robustness of state-of-the-art methods for obtaining CEs. Since a lack of robustness may compromise the validity of CEs, techniques to mitigate this risk are in order. In this survey, we review works in the rapidly growing area of robust CEs and perform an in-depth analysis of the forms of robustness they consider. We also discuss existing solutions and their limitations, providing a solid foundation for future developments.
With large language models, robots can understand language more flexibly and more capable than ever before. This survey reviews and situates recent literature into a spectrum with two poles: 1) mapping between language and some manually defined formal representation of meaning, and 2) mapping between language and high-dimensional vector spaces that translate directly to low-level robot policy. Using a formal representation allows the meaning of the language to be precisely represented, limits the size of the learning problem, and leads to a framework for interpretability and formal safety guarantees. Methods that embed language and perceptual data into high-dimensional spaces avoid this manually specified symbolic structure and thus have the potential to be more general when fed enough data but require more data and computing to train. We discuss the benefits and tradeoffs of each approach and finish by providing directions for future work that achieves the best of both worlds.
Reassessing Evaluation Functions in Algorithmic Recourse: An Empirical Study from a Human-Centered Perspective
PDF ↗In this study, we critically examine the foundational premise of algorithmic recourse - a process of generating counterfactual action plans (i.e., recourses) assisting individuals to reverse adverse decisions made by AI systems. The assumption underlying algorithmic recourse is that individuals accept and act on recourses that minimize the gap between their current and desired states. This assumption, however, remains empirically unverified. To address this issue, we conducted a user study with 362 participants and assessed whether minimizing the distance function, a metric of the gap between the current and desired states, indeed prompts them to accept and act upon suggested recourses. Our findings reveal a nuanced landscape: participants' acceptance of recourses did not correlate with the recourse distance. Moreover, participants' willingness to act upon recourses peaked at the minimal recourse distance but was otherwise constant. These findings cast doubt on the prevailing assumption of algorithmic recourse research and signal the need to rethink the evaluation functions to pave the way for human-centered recourse generation.
In the evolving landscape of human-centered AI, fostering a synergistic relationship between humans and AI agents in decision-making processes stands as a paramount challenge. This work considers a problem setup where an intelligent agent comprising a neural network-based prediction component and a deep reinforcement learning component provides advice to a human decision-maker in complex repeated decision-making environments. Whether the human decision-maker would follow the agent's advice depends on their beliefs and trust in the agent and on their understanding of the advice itself. To this end, we developed an approach named ADESSE to generate explanations about the adviser agent to improve human trust and decision-making. Computational experiments on a range of environments with varying model sizes demonstrate the applicability and scalability of ADESSE. Furthermore, an interactive game-based user study shows that participants were significantly more satisfied, achieved a higher reward in the game, and took less time to select an action when presented with explanations generated by ADESSE. These findings illuminate the critical role of tailored, human-centered explanations in AI-assisted decision-making.
Explaining the singability of lyrics is an important but missing ability of language models (LMs) in song lyrics generation. This ability allows songwriters to quickly assess if LM-generated lyrics can be sung harmoniously with melodies and helps singers align lyrics with melodies during practice. This paper presents XAI-Lyricist, leveraging musical prosody to guide LMs in generating singable lyrics and providing human-understandable singability explanations. We employ a Transformer model to generate lyrics under musical prosody constraints and provide demonstrations of the lyrics' prosody patterns as singability explanations. XAI-Lyricist is evaluated by computational metrics (perplexity, prosody-BLEU) and a human-grounded study (human ratings, average time and number of attempts for singing). Experimental results show that musical prosody can significantly improve the singability of LM-generated lyrics. A controlled study with 14 singers also confirms the usefulness of the provided explanations in helping them to interpret lyrical singability faster than reading plain text lyrics.
It is known that when interacting with explainable autonomous systems, user characteristics are important in determining the most appropriate explanation, but understanding which user characteristics are most relevant to consider is not simple. This paper explores such characteristics and analyses how they affect the perceived usefulness of four types of explanations based on the robot's mental states. These types are belief, goal, hybrid (goal and belief) and baseline explanations. In this study, the explanations were evaluated in the context of a domestic service robot. The user characteristics considered are the perception of the robot's rationality and autonomy, the acceptance of the robot and the user's cognitive tendencies. We found differences in perceived usefulness between explanation types based on user characteristics, with hybrid explanations being the most useful.
Are They the Same Picture? Adapting Concept Bottleneck Models for Human-AI Collaboration in Image Retrieval
PDF ↗Image retrieval plays a pivotal role in applications from wildlife conservation to healthcare, for finding individual animals or relevant images to aid diagnosis. Although deep learning techniques for image retrieval have advanced significantly, their imperfect real-world performance often necessitates including human expertise. Human-in-the-loop approaches typically rely on humans completing the task independently and then combining their opinions with an AI model in various ways, as these models offer very little interpretability or correctability. To allow humans to intervene in the AI model instead, thereby saving human time and effort, we adapt the Concept Bottleneck Model (CBM) and propose CHAIR. CHAIR (a) enables humans to correct intermediate concepts, which helps \textit{improve} embeddings generated, and (b) allows for flexible levels of intervention that accommodate varying levels of human expertise for better retrieval. To show the efficacy of \methodname, we demonstrate that our method performs better than similar models on image retrieval metrics without any external intervention. Furthermore, we also showcase how human intervention helps further improve retrieval performance, thereby achieving human-AI complementarity.
We propose a new graph convolutional block, called MusGConv, specifically designed for the efficient processing of musical score data and motivated by general perceptual principles. It focuses on two fundamental dimensions of music, pitch and rhythm, and considers both relative and absolute representations of these components. We evaluate our approach on four different musical understanding problems: monophonic voice separation, harmonic analysis, cadence detection, and composer identification which, in abstract terms, translate to different graph learning problems, namely, node classification, link prediction, and graph classification. Our experiments demonstrate that MusGConv improves the performance on three of the aforementioned tasks while being conceptually very simple and efficient. We interpret this as evidence that it is beneficial to include perception-informed processing of fundamental musical concepts when developing graph network applications on musical score data. All code and models are released on https://github.com/manoskary/musgconv.
Exploring the narratives conveyed by fine-art paintings is a challenge in image captioning, where the goal is to generate descriptions that not only precisely represent the visual content but also offer a in-depth interpretation of the artwork's meaning. The task is particularly complex for artwork images due to their diverse interpretations and varied aesthetic principles across different artistic schools and styles. In response to this, we present KALE (Knowledge-Augmented vision-Language model for artwork Elaborations), a novel approach that enhances existing vision-language models by integrating artwork metadata as additional knowledge. KALE incorporates the metadata in two ways: firstly as direct textual input, and secondly through a multimodal heterogeneous knowledge graph. To optimize the learning of graph representations, we introduce a new cross-modal alignment loss that maximizes the similarity between the image and its corresponding metadata. Experimental results demonstrate that KALE achieves strong performance (when evaluated with CIDEr, in particular) over existing state-of-the-art work across several artwork datasets. Source code of the project is available at https://github.com/Yanbei-Jiang/Artwork-Interpretation.
VulnerabilityMap: An Open Framework for Mapping Vulnerability among Urban Disadvantaged Populations in the United States
PDF ↗Cities are crucibles of numerous opportunities, but also hotbeds of inequality. The plight of disadvantaged populations who are ``left behind'' within urban environments has been an increasingly pressing concern, which poses substantial threats to the realization of the UN SDG agenda. However, a comprehensive framework for studying this urban dilemma is currently absent, preventing researchers from developing AI models for social good prediction and intervention. To fill this gap, we construct VulnerabilityMap, a framework to meticulously dissect the challenges faced by urban disadvantaged populations, unraveling their vulnerability to a spectrum of shocks and stresses that are categorized through the prism of Maslow's hierarchy of needs. Specifically, we systematically collect large-scale multi-sourced census and web-based data covering more than 328 million people in the United States regarding demographic features, neighborhood environments, offline mobility behaviors, and online social connections. These features are further related to vulnerability outcomes from short-term shocks such as COVID-19 and long-term physiological, social, and self-actualization stresses. Leveraging our framework, we construct machine learning models that exhibit strong performance in predicting vulnerability outcomes from various disadvantage features, which shows the promising utility of our framework to support targeted AI models. Moreover, we provide model-based explainability analysis to interpret the reasons underlying model predictions, shedding light on intricate social factors that trap certain populations inside vulnerable situations. Our constructed dataset is publicly available at https://github.com/LinChen-65/VulnerabilityMap/.
Functional Graph Convolutional Networks: A Unified Multi-task and Multi-modal Learning Framework to Facilitate Health and Social-Care Insights
PDF ↗This paper introduces a novel Functional Graph Convolutional Network (funGCN) framework that combines Functional Data Analysis and Graph Convolutional Networks to address the complexities of multi-task and multi-modal learning in digital health and longitudinal studies. With the growing importance of health solutions to improve health care and social support, ensure healthy lives, and promote well-being at all ages, funGCN offers a unified approach to handle multivariate longitudinal data for multiple entities and ensures interpretability even with small sample sizes. Key innovations include task-specific embedding components that manage different data types, the ability to perform classification, regression, and forecasting, and the creation of a knowledge graph for insightful data interpretation. The efficacy of funGCN is validated through simulation experiments and a real-data application. funGCN source code is publicly available at https://github.com/IBM/funGCN.
Integrating Intent Understanding and Optimal Behavior Planning for Behavior Tree Generation from Human Instructions
PDF ↗Robots executing tasks following human instructions in domestic or industrial environments essentially require both adaptability and reliability. Behavior Tree (BT) emerges as an appropriate control architecture for these scenarios due to its modularity and reactivity. Existing BT generation methods, however, either do not involve interpreting natural language or cannot theoretically guarantee the BTs' success. This paper proposes a two-stage framework for BT generation, which first employs large language models (LLMs) to interpret goals from high-level instructions, then constructs an efficient goal-specific BT through the Optimal Behavior Tree Expansion Algorithm (OBTEA). We represent goals as well-formed formulas in first-order logic, effectively bridging intent understanding and optimal behavior planning. Experiments in the service robot validate the proficiency of LLMs in producing grammatically correct and accurately interpreted goals, demonstrate OBTEA's superiority over the baseline BT Expansion algorithm in various metrics, and finally confirm the practical deployability of our framework. The project website is https://dids-ei.github.io/Project/LLM-OBTEA.
Vision-fused Attack: Advancing Aggressive and Stealthy Adversarial Text against Neural Machine Translation
PDF ↗While neural machine translation (NMT) models achieve success in our daily lives, they show vulnerability to adversarial attacks. Despite being harmful, these attacks also offer benefits for interpreting and enhancing NMT models, thus drawing increased research attention. However, existing studies on adversarial attacks are insufficient in both attacking ability and human imperceptibility due to their sole focus on the scope of language. This paper proposes a novel vision-fused attack (VFA) framework to acquire powerful adversarial text, i.e., more aggressive and stealthy. Regarding the attacking ability, we design the vision-merged solution space enhancement strategy to enlarge the limited semantic solution space, which enables us to search for adversarial candidates with higher attacking ability. For human imperceptibility, we propose the perception-retained adversarial text selection strategy to align the human text-reading mechanism. Thus, the finally selected adversarial text could be more deceptive. Extensive experiments on various models, including large language models (LLMs) like LLaMA and GPT-3.5, strongly support that VFA outperforms the comparisons by large margins (up to 81%/14% improvements on ASR/SSIM).
It Ain’t That Bad: Understanding the Mysterious Performance Drop in OOD Generalization for Generative Transformer Models
PDF ↗Large language models (LLMs) have achieved remarkable proficiency on solving diverse problems. However, their generalization ability is not always satisfying and the generalization problem is common for generative transformer models in general. Researchers take basic mathematical tasks like n-digit addition or multiplication as important perspectives for investigating their generalization behaviors. It is observed that when training models on n-digit operations (e.g., additions) in which both input operands are n-digit in length, models generalize successfully on unseen n-digit inputs (in-distribution (ID) generalization), but fail miserably on longer, unseen cases (out-of-distribution (OOD) generalization). We bring this unexplained performance drop into attention and ask whether there is systematic OOD generalization. Towards understanding LLMs, we train various smaller language models which may share the same underlying mechanism. We discover that the strong ID generalization stems from structured representations, while behind the unsatisfying OOD performance, the models still exhibit clear learned algebraic structures. Specifically, these models map unseen OOD inputs to outputs with learned equivalence relations in the ID domain, which we call the equivalence generalization. These findings deepen our knowledge regarding the generalizability of generative models including LLMs, and provide insights into potential avenues for improvement.
The current approaches for fake news detection are mainly devoted to extracting candidate evidence from comments (or external articles) and establishing interactive reasoning with the news itself to verify the falsehood of the news. However, they still have several drawbacks: 1) The interaction object is coarse-grained, which mainly drives the entire news to participate in interaction, but ignores the learning of potential suspicious segments in news; 2) The reasoning ways are relatively single, making it difficult to explore the various possible correlations between news and candidate evidence. To this end, we propose Unified Evidence Enhancement Inference framework (UEEI) to discover and infer high-quality evidence to reveal the false parts of news for detection. Specifically, UEEI first promotes the interaction fusion between comments and news from the perspectives of semantic and emotion, thereby learning potential suspicious fragments in news. Then, the model constructs entity-level and relationship-level retrievals to screen sufficient candidate evidence from external sources. Finally, we measure coherence between suspicious fragments and candidate evidence by multi-view reasoning, and further infer explainable evidence that discovers the false parts of news. Experiments on three public datasets confirm the effectiveness and interpretability of our UEEI.
Knowledge distillation, transferring knowledge from a teacher model to a student model, has emerged as a powerful technique in neural machine translation for compressing models or simplifying training targets. Knowledge distillation encompasses two primary methods: sentence-level distillation and token-level distillation. In sentence-level distillation, the student model is trained to align with the output of the teacher model, which can alleviate the training difficulty and give student model a comprehensive understanding of global structure. Differently, token-level distillation requires the student model to learn the output distribution of the teacher model, facilitating a more fine-grained transfer of knowledge. Studies have revealed divergent performances between sentence-level and token-level distillation across different scenarios, leading to the confusion on the empirical selection of knowledge distillation methods. In this study, we argue that token-level distillation, with its more complex objective (i.e., distribution), is better suited for “simple” scenarios, while sentence-level distillation excels in “complex” scenarios. To substantiate our hypothesis, we systematically analyze the performance of distillation methods by varying the model size of student models, the complexity of text, and the difficulty of decoding procedure. While our experimental results validate our hypothesis, defining the complexity level of a given scenario remains a challenging task. So we further introduce a novel hybrid method that combines token-level and sentence-level distillation through a gating mechanism, aiming to leverage the advantages of both individual methods. Experiments demonstrate that the hybrid method surpasses the performance of token-level or sentence-level distillation methods and the previous works by a margin, demonstrating the effectiveness of the proposed hybrid method.
Training data plays a pivotal role in AI models. Large language models (LLMs) are trained with massive amounts of documents, and their parameters hold document-related contents. Recently, several studies identified content-specific locations in LLMs by examining the parameters. Instead of the post hoc interpretation, we propose another approach. We propose document-wise memory architecture to track document memories in training. The proposed architecture maps document representations to memory entries, which softly mask memories in the forward process of LLMs. Additionally, we propose document guidance loss, which increases the likelihood of text with document memories and reduces the likelihood of the text with the memories of other documents. Experimental results on Wikitext-103-v1 with Pythia-1B show that the proposed methods provide different memory entries for documents and high recall of document-related content in generation with trained document-wise memories.