Research Papers

Interpretability at Scale: Identifying Causal Mechanisms in Alpaca

Zhengxuan Wu, Atticus Geiger, Thomas Icard +2

Obtaining human-interpretable explanations of large, general-purpose language models is an urgent goal for AI safety. However, it is just as important that our interpretability methods are faithful to the causal dynamics underlying model behavior and able to robustly generalize to unseen inputs. Distributed Alignment Search (DAS) is a powerful gradient descent method grounded in a theory of causal abstraction that has uncovered perfect alignments between interpretable symbolic algorithms and small deep learning models fine-tuned for specific tasks. In the present paper, we scale DAS significantly by replacing the remaining brute-force search steps with learned parameters -- an approach we call Boundless DAS. This enables us to efficiently search for interpretable causal structure in large language models while they follow instructions. We apply Boundless DAS to the Alpaca model (7B parameters), which, off the shelf, solves a simple numerical reasoning problem. With Boundless DAS, we discover that Alpaca does this by implementing a causal model with two interpretable boolean variables. Furthermore, we find that the alignment of neural representations with these variables is robust to changes in inputs and instructions. These findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models. Our tool is extensible to larger LLMs and is released publicly at `https://github.com/stanfordnlp/pyvene`.

How can I choose an explainer? An Application-grounded Evaluation of Post-hoc Explanations

Sérgio Jesus, Catarina Belém, Vladimir Balayan +4

There have been several research works proposing new Explainable AI (XAI) methods designed to generate model explanations having specific properties, or desiderata, such as fidelity, robustness, or human-interpretability. However, explanations are seldom evaluated based on their true practical impact on decision-making tasks. Without that assessment, explanations might be chosen that, in fact, hurt the overall performance of the combined system of ML model + end-users. This study aims to bridge this gap by proposing XAI Test, an application-grounded evaluation methodology tailored to isolate the impact of providing the end-user with different levels of information. We conducted an experiment following XAI Test to evaluate three popular post-hoc explanation methods -- LIME, SHAP, and TreeInterpreter -- on a real-world fraud detection task, with real data, a deployed ML model, and fraud analysts. During the experiment, we gradually increased the information provided to the fraud analysts in three stages: Data Only, i.e., just transaction data without access to model score nor explanations, Data + ML Model Score, and Data + ML Model Score + Explanations. Using strong statistical analysis, we show that, in general, these popular explainers have a worse impact than desired. Some of the conclusion highlights include: i) showing Data Only results in the highest decision accuracy and the slowest decision time among all variants tested, ii) all the explainers improve accuracy over the Data + ML Model Score variant but still result in lower accuracy when compared with Data Only; iii) LIME was the least preferred by users, probably due to its substantially lower variability of explanations from case to case.

When to Make Exceptions: Exploring Language Models as Accounts of Human Moral Judgment

Zhijing Jin, Sydney Levine, Fernando Gonzalez +6

AI systems are becoming increasingly intertwined with human life. In order to effectively collaborate with humans and ensure safety, AI systems need to be able to understand, interpret and predict human moral judgments and decisions. Human moral judgments are often guided by rules, but not always. A central challenge for AI safety is capturing the flexibility of the human moral mind -- the ability to determine when a rule should be broken, especially in novel or unusual situations. In this paper, we present a novel challenge set consisting of rule-breaking question answering (RBQA) of cases that involve potentially permissible rule-breaking -- inspired by recent moral psychology studies. Using a state-of-the-art large language model (LLM) as a basis, we propose a novel moral chain of thought (MORALCOT) prompting strategy that combines the strengths of LLMs with theories of moral reasoning developed in cognitive science to predict human moral judgments. MORALCOT outperforms seven existing LLMs by 6.2% F1, suggesting that modeling human reasoning might be necessary to capture the flexibility of the human moral mind. We also conduct a detailed error analysis to suggest directions for future work to improve AI safety using RBQA. Our data is open-sourced at https://huggingface.co/datasets/feradauto/MoralExceptQA and code at https://github.com/feradauto/MoralCoT

Beyond Explaining: Opportunities and Challenges of XAI-Based Model Improvement

Leander Weber, Sebastian Lapuschkin, Alexander Binder +1

Explainable Artificial Intelligence (XAI) is an emerging research field bringing transparency to highly complex and opaque machine learning (ML) models. Despite the development of a multitude of methods to explain the decisions of black-box classifiers in recent years, these tools are seldomly used beyond visualization purposes. Only recently, researchers have started to employ explanations in practice to actually improve models. This paper offers a comprehensive overview over techniques that apply XAI practically for improving various properties of ML models, and systematically categorizes these approaches, comparing their respective strengths and weaknesses. We provide a theoretical perspective on these methods, and show empirically through experiments on toy and realistic settings how explanations can help improve properties such as model generalization ability or reasoning, among others. We further discuss potential caveats and drawbacks of these methods. We conclude that while model improvement based on XAI can have significant beneficial effects even on complex and not easily quantifyable model properties, these methods need to be applied carefully, since their success can vary depending on a multitude of factors, such as the model and dataset used, or the employed explanation method.

Machine learning systems have become popular in fields such as marketing, financing, or data mining. While they are highly accurate, complex machine learning systems pose challenges for engineers and users. Their inherent complexity makes it impossible to easily judge their fairness and the correctness of statistically learned relations between variables and classes. Explainable AI aims to solve this challenge by modelling explanations alongside with the classifiers, potentially improving user trust and acceptance. However, users should not be fooled by persuasive, yet untruthful explanations. We therefore conduct a user study in which we investigate the effects of model accuracy and explanation fidelity, i.e. how truthfully the explanation represents the underlying model, on user trust. Our findings show that accuracy is more important for user trust than explainability. Adding an explanation for a classification result can potentially harm trust, e.g. when adding nonsensical explanations. We also found that users cannot be tricked by high-fidelity explanations into having trust for a bad classifier. Furthermore, we found a mismatch between observed (implicit) and self-reported (explicit) trust.

Do Concept Bottleneck Models Learn as Intended?

Andrei Margeloiu, Matthew Ashman, Umang Bhatt +3

Concept bottleneck models map from raw inputs to concepts, and then from concepts to targets. Such models aim to incorporate pre-specified, high-level concepts into the learning procedure, and have been motivated to meet three desiderata: interpretability, predictability, and intervenability. However, we find that concept bottleneck models struggle to meet these goals. Using post hoc interpretability methods, we demonstrate that concepts do not correspond to anything semantically meaningful in input space, thus calling into question the usefulness of concept bottleneck models in their current form.

Evaluating Standard Feature Sets Towards Increased Generalisability and Explainability of ML-based Network Intrusion Detection

Mohanad Sarhan, Siamak Layeghy, Marius Portmann

Machine Learning (ML)-based network intrusion detection systems bring many benefits for enhancing the cybersecurity posture of an organisation. Many systems have been designed and developed in the research community, often achieving a close to perfect detection rate when evaluated using synthetic datasets. However, the high number of academic research has not often translated into practical deployments. There are several causes contributing towards the wide gap between research and production, such as the limited ability of comprehensive evaluation of ML models and lack of understanding of internal ML operations. This paper tightens the gap by evaluating the generalisability of a common feature set to different network environments and attack scenarios. Therefore, two feature sets (NetFlow and CICFlowMeter) have been evaluated in terms of detection accuracy across three key datasets, i.e., CSE-CIC-IDS2018, BoT-IoT, and ToN-IoT. The results show the superiority of the NetFlow feature set in enhancing the ML models detection accuracy of various network attacks. In addition, due to the complexity of the learning models, SHapley Additive exPlanations (SHAP), an explainable AI methodology, has been adopted to explain and interpret the classification decisions of ML models. The Shapley values of two common feature sets have been analysed across multiple datasets to determine the influence contributed by each feature towards the final ML prediction.

The lack of explainability of a decision from an Artificial Intelligence (AI) based "black box" system/model, despite its superiority in many real-world applications, is a key stumbling block for adopting AI in many high stakes applications of different domain or industry. While many popular Explainable Artificial Intelligence (XAI) methods or approaches are available to facilitate a human-friendly explanation of the decision, each has its own merits and demerits, with a plethora of open challenges. We demonstrate popular XAI methods with a mutual case study/task (i.e., credit default prediction), analyze for competitive advantages from multiple perspectives (e.g., local, global), provide meaningful insight on quantifying explainability, and recommend paths towards responsible or human-centered AI using XAI as a medium. Practitioners can use this work as a catalog to understand, compare, and correlate competitive advantages of popular XAI methods. In addition, this survey elicits future research directions towards responsible or human-centric AI systems, which is crucial to adopt AI in high stakes applications.

With the advances in artificial intelligence (AI), data-driven algorithms are becoming increasingly popular in the medical domain. However, due to the nonlinear and complex behavior of many of these algorithms, decision-making by such algorithms is not trustworthy for clinicians and is considered a black-box process. Hence, the scientific community has introduced explainable artificial intelligence (XAI) to remedy the problem. This systematic scoping review investigates the application of XAI in breast cancer detection and risk prediction. We conducted a comprehensive search on Scopus, IEEE Explore, PubMed, and Google Scholar (first 50 citations) using a systematic search strategy. The search spanned from January 2017 to July 2023, focusing on peer-reviewed studies implementing XAI methods in breast cancer datasets. Thirty studies met our inclusion criteria and were included in the analysis. The results revealed that SHapley Additive exPlanations (SHAP) is the top model-agnostic XAI technique in breast cancer research in terms of usage, explaining the model prediction results, diagnosis and classification of biomarkers, and prognosis and survival analysis. Additionally, the SHAP model primarily explained tree-based ensemble machine learning models. The most common reason is that SHAP is model agnostic, which makes it both popular and useful for explaining any model prediction. Additionally, it is relatively easy to implement effectively and completely suits performant models, such as tree-based models. Explainable AI improves the transparency, interpretability, fairness, and trustworthiness of AI-enabled health systems and medical devices and, ultimately, the quality of care and outcomes.

Artificial Intelligence (AI) has created the single biggest technology revolution the world has ever seen. For the finance sector, it provides great opportunities to enhance customer experience, democratize financial services, ensure consumer protection and significantly improve risk management. While it is easier than ever to run state-of-the-art machine learning models, designing and implementing systems that support real-world finance applications have been challenging. In large part because they lack transparency and explainability which are important factors in establishing reliable technology and the research on this topic with a specific focus on applications in credit risk management. In this paper, we implement two advanced post-hoc model agnostic explainability techniques called Local Interpretable Model Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) to machine learning (ML)-based credit scoring models applied to the open-access data set offered by the US-based P2P Lending Platform, Lending Club. Specifically, we use LIME to explain instances locally and SHAP to get both local and global explanations. We discuss the results in detail and present multiple comparison scenarios by using various kernels available for explaining graphs generated using SHAP values. We also discuss the practical challenges associated with the implementation of these state-of-art eXplainabale AI (XAI) methods and document them for future reference. We have made an effort to document every technical aspect of this research, while at the same time providing a general summary of the conclusions.

Relating the Partial Dependence Plot and Permutation Feature Importance to the Data Generating Process

Christoph Molnar, Timo Freiesleben, Gunnar König +3

Scientists and practitioners increasingly rely on machine learning to model data and draw conclusions. Compared to statistical modeling approaches, machine learning makes fewer explicit assumptions about data structures, such as linearity. However, their model parameters usually cannot be easily related to the data generating process. To learn about the modeled relationships, partial dependence (PD) plots and permutation feature importance (PFI) are often used as interpretation methods. However, PD and PFI lack a theory that relates them to the data generating process. We formalize PD and PFI as statistical estimators of ground truth estimands rooted in the data generating process. We show that PD and PFI estimates deviate from this ground truth due to statistical biases, model variance and Monte Carlo approximation errors. To account for model variance in PD and PFI estimation, we propose the learner-PD and the learner-PFI based on model refits, and propose corrected variance and confidence interval estimators.

Explainable AI: current status and future directions

Prashant Gohel, Priyanka Singh, Manoranjan Mohanty

Explainable Artificial Intelligence (XAI) is an emerging area of research in the field of Artificial Intelligence (AI). XAI can explain how AI obtained a particular solution (e.g., classification or object detection) and can also answer other "wh" questions. This explainability is not possible in traditional AI. Explainability is essential for critical applications, such as defense, health care, law and order, and autonomous driving vehicles, etc, where the know-how is required for trust and transparency. A number of XAI techniques so far have been purposed for such applications. This paper provides an overview of these techniques from a multimedia (i.e., text, image, audio, and video) point of view. The advantages and shortcomings of these techniques have been discussed, and pointers to some future directions have also been provided.