Research Papers

Necessity and sufficiency are the building blocks of all successful explanations. Yet despite their importance, these notions have been conceptually underdeveloped and inconsistently applied in explainable artificial intelligence (XAI), a fast-growing research area that is so far lacking in firm theoretical foundations. Building on work in logic, probability, and causality, we establish the central role of necessity and sufficiency in XAI, unifying seemingly disparate methods in a single formal framework. We provide a sound and complete algorithm for computing explanatory factors with respect to a given context, and demonstrate its flexibility and competitive performance against state of the art alternatives on various tasks.

Software defect prediction models are classifiers that are constructed from historical software data. Such software defect prediction models have been proposed to help developers optimize the limited Software Quality Assurance (SQA) resources and help managers develop SQA plans. Prior studies have different goals for their defect prediction models and use different techniques for generating visual explanations of their models. Yet, it is unclear what are the practitioners' perceptions of (1) these defect prediction model goals, and (2) the model-agnostic techniques used to visualize these models. We conducted a qualitative survey to investigate practitioners' perceptions of the goals of defect prediction models and the model-agnostic techniques used to generate visual explanations of defect prediction models. We found that (1) 82%-84% of the respondents perceived that the three goals of defect prediction models are useful; (2) LIME is the most preferred technique for understanding the most important characteristics that contributed to a prediction of a file, while ANOVA/VarImp is the second most preferred technique for understanding the characteristics that are associated with software defects in the past. Our findings highlight the significance of investigating how to improve the understanding of defect prediction models and their predictions. Hence, model-agnostic techniques from explainable AI domain may help practitioners to understand defect prediction models and their predictions.

Artificial intelligence (AI) is a technology which is increasingly being utilised in society and the economy worldwide, and its implementation is planned to become more prevalent in coming years. AI is increasingly being embedded in our lives, supplementing our pervasive use of digital technologies. But this is being accompanied by disquiet over problematic and dangerous implementations of AI, or indeed, even AI itself deciding to do dangerous and problematic actions, especially in fields such as the military, medicine and criminal justice. These developments have led to concerns about whether and how AI systems adhere, and will adhere to ethical standards. These concerns have stimulated a global conversation on AI ethics, and have resulted in various actors from different countries and sectors issuing ethics and governance initiatives and guidelines for AI. Such developments form the basis for our research in this report, combining our international and interdisciplinary expertise to give an insight into what is happening in Australia, China, Europe, India and the US.

BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models

Yi Zeng, Weiyu Sun, Tran Ngoc Huynh +3

Safety backdoor attacks in large language models (LLMs) enable the stealthy triggering of unsafe behaviors while evading detection during normal interactions. The high dimensionality of potential triggers in the token space and the diverse range of malicious behaviors make this a critical challenge. We present BEEAR, a mitigation approach leveraging the insight that backdoor triggers induce relatively uniform drifts in the model's embedding space. Our bi-level optimization method identifies universal embedding perturbations that elicit unwanted behaviors and adjusts the model parameters to reinforce safe behaviors against these perturbations. Experiments show BEEAR reduces the success rate of RLHF time backdoor attacks from >95% to <1% and from 47% to 0% for instruction-tuning time backdoors targeting malicious code generation, without compromising model utility. Requiring only defender-defined safe and unwanted behaviors, BEEAR represents a step towards practical defenses against safety backdoors in LLMs, providing a foundation for further advancements in AI safety and security.

SimpleSafetyTests: a Test Suite for Identifying Critical Safety Risks in Large Language Models

Bertie Vidgen, Nino Scherrer, Hannah Rose Kirk +4

The past year has seen rapid acceleration in the development of large language models (LLMs). However, without proper steering and safeguards, LLMs will readily follow malicious instructions, provide unsafe advice, and generate toxic content. We introduce SimpleSafetyTests (SST) as a new test suite for rapidly and systematically identifying such critical safety risks. The test suite comprises 100 test prompts across five harm areas that LLMs, for the vast majority of applications, should refuse to comply with. We test 11 open-access and open-source LLMs and four closed-source LLMs, and find critical safety weaknesses. While some of the models do not give a single unsafe response, most give unsafe responses to more than 20% of the prompts, with over 50% unsafe responses in the extreme. Prepending a safety-emphasising system prompt substantially reduces the occurrence of unsafe responses, but does not completely stop them from happening. Trained annotators labelled every model response to SST (n = 3,000). We use these annotations to evaluate five AI safety filters (which assess whether a models' response is unsafe given a prompt) as a way of automatically evaluating models' performance on SST. The filters' performance varies considerably. There are also differences across the five harm areas, and on the unsafe versus safe responses. The widely-used Perspective API has 72% accuracy and a newly-created zero-shot prompt to OpenAI's GPT-4 performs best with 89% accuracy. Content Warning: This paper contains prompts and responses that relate to child abuse, suicide, self-harm and eating disorders, scams and fraud, illegal items, and physical harm.

Rethinking Model Evaluation as Narrowing the Socio-Technical Gap

Q. Vera Liao, Ziang Xiao

The recent development of generative large language models (LLMs) poses new challenges for model evaluation that the research community and industry have been grappling with. While the versatile capabilities of these models ignite much excitement, they also inevitably make a leap toward homogenization: powering a wide range of applications with a single, often referred to as ``general-purpose'', model. In this position paper, we argue that model evaluation practices must take on a critical task to cope with the challenges and responsibilities brought by this homogenization: providing valid assessments for whether and how much human needs in diverse downstream use cases can be satisfied by the given model (\textit{socio-technical gap}). By drawing on lessons about improving research realism from the social sciences, human-computer interaction (HCI), and the interdisciplinary field of explainable AI (XAI), we urge the community to develop evaluation methods based on real-world contexts and human requirements, and embrace diverse evaluation methods with an acknowledgment of trade-offs between realisms and pragmatic costs to conduct the evaluation. By mapping HCI and current NLG evaluation methods, we identify opportunities for evaluation methods for LLMs to narrow the socio-technical gap and pose open questions.

Interpreting and explaining the behavior of deep neural networks is critical for many tasks. Explainable AI provides a way to address this challenge, mostly by providing per-pixel relevance to the decision. Yet, interpreting such explanations may require expert knowledge. Some recent attempts toward interpretability adopt a concept-based framework, giving a higher-level relationship between some concepts and model decisions. This paper proposes Bottleneck Concept Learner (BotCL), which represents an image solely by the presence/absence of concepts learned through training over the target task without explicit supervision over the concepts. It uses self-supervision and tailored regularizers so that learned concepts can be human-understandable. Using some image classification tasks as our testbed, we demonstrate BotCL's potential to rebuild neural networks for better interpretability. Code is available at https://github.com/wbw520/BotCL and a simple demo is available at https://botcl.liangzhili.com/.

Adversarial Training for High-Stakes Reliability

Daniel M. Ziegler, Seraphina Nix, Lawrence Chan +9

In the future, powerful AI systems may be deployed in high-stakes settings, where a single failure could be catastrophic. One technique for improving AI safety in high-stakes settings is adversarial training, which uses an adversary to generate examples to train on in order to achieve better worst-case performance. In this work, we used a safe language generation task (``avoid injuries'') as a testbed for achieving high reliability through adversarial training. We created a series of adversarial training techniques -- including a tool that assists human adversaries -- to find and eliminate failures in a classifier that filters text completions suggested by a generator. In our task, we determined that we can set very conservative classifier thresholds without significantly impacting the quality of the filtered outputs. We found that adversarial training increased robustness to the adversarial attacks that we trained on -- doubling the time for our contractors to find adversarial examples both with our tool (from 13 to 26 minutes) and without (from 20 to 44 minutes) -- without affecting in-distribution performance. We hope to see further work in the high-stakes reliability setting, including more powerful tools for enhancing human adversaries and better ways to measure high levels of reliability, until we can confidently rule out the possibility of catastrophic deployment-time failures of powerful models.

This paper presents an empirical study on the weights of neural networks, where we interpret each model as a point in a high-dimensional space -- the neural weight space. To explore the complex structure of this space, we sample from a diverse selection of training variations (dataset, optimization procedure, architecture, etc.) of neural network classifiers, and train a large number of models to represent the weight space. Then, we use a machine learning approach for analyzing and extracting information from this space. Most centrally, we train a number of novel deep meta-classifiers with the objective of classifying different properties of the training setup by identifying their footprints in the weight space. Thus, the meta-classifiers probe for patterns induced by hyper-parameters, so that we can quantify how much, where, and when these are encoded through the optimization process. This provides a novel and complementary view for explainable AI, and we show how meta-classifiers can reveal a great deal of information about the training setup and optimization, by only considering a small subset of randomly selected consecutive weights. To promote further research on the weight space, we release the neural weight space (NWS) dataset -- a collection of 320K weight snapshots from 16K individually trained deep neural networks.

Artificial intelligence (AI) powered wireless networks promise to revolutionize the conventional operation and structure of current networks from network design to infrastructure management, cost reduction, and user performance improvement. Empowering future networks with AI functionalities will enable a shift from reactive/incident driven operations to proactive/data-driven operations. This paper provides an overview on the integration of AI functionalities in 5G and beyond networks. Key factors for successful AI integration such as data, security, and explainable AI are highlighted. We also summarize the various types of network intelligence as well as machine learning based air interface in future networks. Use case examples for the application of AI to the wireless domain are then summarized. We highlight on applications to the physical layer, mobility management, wireless security, and localization.

Being able to explain the prediction to clinical end-users is a necessity to leverage the power of artificial intelligence (AI) models for clinical decision support. For medical images, a feature attribution map, or heatmap, is the most common form of explanation that highlights important features for AI models' prediction. However, it is unknown how well heatmaps perform on explaining decisions on multi-modal medical images, where each image modality or channel visualizes distinct clinical information of the same underlying biomedical phenomenon. Understanding such modality-dependent features is essential for clinical users' interpretation of AI decisions. To tackle this clinically important but technically ignored problem, we propose the modality-specific feature importance (MSFI) metric. It encodes clinical image and explanation interpretation patterns of modality prioritization and modality-specific feature localization. We conduct a clinical requirement-grounded, systematic evaluation using computational methods and a clinician user study. Results show that the examined 16 heatmap algorithms failed to fulfill clinical requirements to correctly indicate AI model decision process or decision quality. The evaluation and MSFI metric can guide the design and selection of XAI algorithms to meet clinical requirements on multi-modal explanation.

Innovations in data science and AI/ML have a central role to play in supporting global efforts to combat COVID-19. The versatility of AI/ML technologies enables scientists and technologists to address an impressively broad range of biomedical, epidemiological, and socioeconomic challenges. This wide-reaching scientific capacity, however, also raises a diverse array of ethical challenges. The need for researchers to act quickly and globally in tackling SARS-CoV-2 demands unprecedented practices of open research and responsible data sharing at a time when innovation ecosystems are hobbled by proprietary protectionism, inequality, and a lack of public trust. Moreover, societally impactful interventions like digital contact tracing are raising fears of surveillance creep and are challenging widely held commitments to privacy, autonomy, and civil liberties. Prepandemic concerns that data-driven innovations may function to reinforce entrenched dynamics of societal inequity have likewise intensified given the disparate impact of the virus on vulnerable social groups and the life-and-death consequences of biased and discriminatory public health outcomes. To address these concerns, I offer five steps that need to be taken to encourage responsible research and innovation. These provide a practice-based path to responsible AI/ML design and discovery centered on open, accountable, equitable, and democratically governed processes and products. When taken from the start, these steps will not only enhance the capacity of innovators to tackle COVID-19 responsibly, they will, more broadly, help to better equip the data science and AI/ML community to cope with future pandemics and to support a more humane, rational, and just society.