Conditional diffusion models have the generative controllability by incorporating external conditions. However, their performance significantly degrades with noisy conditions, such as corrupted labels in the image generation or unreliable observations or states in the control policy generation. This paper introduces a robust learning framework to address extremely noisy conditions in conditional diffusion models. We empirically demonstrate that existing noise-robust methods fail when the noise level is high. To overcome this, we propose learning pseudo conditions as surrogates for clean conditions and refining pseudo ones progressively via the technique of temporal ensembling. Additionally, we develop a Reverse-time Diffusion Condition (RDC) technique, which diffuses pseudo conditions to reinforce the memorization effect and further facilitate the refinement of the pseudo conditions. Experimentally, our approach achieves state-of-the-art performance across a range of noise levels on both class-conditional image generation and visuomotor policy generation tasks.The code can be accessible via the project page https://robustdiffusionpolicy.github.io

We introduce a condition monitoring (CM) method that relies on spectral analysis of bearing vibration signals. Assuming that the signal is divided into non-overlapping segments, we design an automatic procedure to compare the estimated log-spectrum of a given segment with the log-spectra of past segments. The differences between the log-spectra of past segments and the current segment are stored as the columns of a matrix, which is factorized using dictionary learning (DL). The resulting matrix factors are the dictionary and the representation matrix. The size of the dictionary and the sparsity level of the representation matrix are selected based on an information theoretic criterion (ITC). We show that this factorization generalizes the factorization derived from the definition of the cepstrum. We define three new CM indicators: the ITC value, the estimated sparsity value, and the distance between the dictionary whose estimation involves the current segment and the dictionary estimated from the initial segments of the signal. We investigate the impact of the region of the frequency domain considered in the comparison and the norm used in the DL objective function on these indicators. The performance of the novel method is demonstrated through experiments conducted on three publicly available datasets representing various conditions: abrupt changes, gradual changes, and a combination of abrupt and gradual changes.

Developing benchmark datasets to tackle the bias problem in large language models (LLMs) is difficult for mixed-ethnic, small, and/or indigenous societies with limited resources. Existing bias benchmark datasets reflect the societal makeup of resource-rich societies such as the US and Europe. A deficit in available annotated datasets, the lack of annotators, and relevant LLM-generated text limit the potential for research in developing debiasing techniques for resource-restricted settings. Practices such as discarding data instances with annotator disagreement or obtaining a majority label from many annotators with multiple iterations of annotations are not applicable in this setting because it could lead to discrimination. Rather than discarding the information from such annotations, we propose utilising annotator disagreement information through a multi-annotator ensemble approach to build bias benchmark datasets. We capture annotator information by obtaining soft labels, which provide probability distributions over the hard labels that are either manually annotated or from pre-trained models. Firstly, we use pre-trained language models as an alternative for scenarios where manual annotations are restricted and demonstrate such readily accessible models yield similar or better performance than baseline aggregated manual annotator labels. Secondly, we demonstrate that classifications using the multi-annotator ensemble approach perform better than the single-label trained classification model.

Forecasting surges in hospital admissions caused by severe respiratory infections is of crucial importance during the winter season to enable proactive hospital management and timely decision-making to prevent healthcare system overload. As time series derived from hospital surveillance systems for these severe cases are sparse and encode weak seasonality patterns, machine learning is key to computing accurate forecasts. The most recent algorithmic advance in time series forecasting is the adaptation of generative pre-trained transformers (GPTs). Those models, pre-trained on large datasets, have the potential to transfer knowledge to smaller datasets, such as hospital surveillance data, for very specific and well-defined case definitions. We demonstrate that despite this great potential for such practical applications, one of the largest first-generation GPTs is not able to provide accurate forecasts and is outperformed by simple linear models and even naïve forecasts.

Recent research has demonstrated the effectiveness of knowledge distillation in Domain Generalization. However, existing approaches often overlook domain-specific knowledge and rely on an offline distillation strategy, limiting the effectiveness of knowledge transfer. To address these limitations, we propose Balanced Online knowLedge Distillation (BOLD). BOLD leverages a multi-domain expert teacher model, with each expert specializing in a specific source domain, enabling the student to distill both domain-invariant and domain-specific knowledge. We incorporate the Pareto optimization principle and uncertainty weighting to balance these two types of knowledge, ensuring simultaneous optimization without compromising either. Additionally, BOLD employs an online knowledge distillation strategy, allowing the teacher and student to learn concurrently. This dynamic interaction enables the teacher to adapt based on student feedback, facilitating more effective knowledge transfer. Extensive experiments on seven benchmarks demonstrate that BOLD outperforms state-of-the-art methods. Furthermore, we provide theoretical insights that highlight the importance of domain-specific knowledge and the advantages of uncertainty weighting.

Causal world models are systems that can answer counterfactual questions about an environment of interest, i.e. predict how it would have evolved if an arbitrary subset of events had been realized differently. It requires understanding the underlying causes behind chains of events and conducting causal inference for arbitrary unseen distributions. So far, this task eludes foundation models, notably large language models (LLMs), which do not have demonstrated causal reasoning capabilities beyond the memorization of existing causal relationships. Furthermore, evaluating counterfactuals in real-world applications is challenging since only the factual world is observed, limiting evaluation to synthetic datasets. We address these problems by explicitly extracting and modeling causal relationships and propose the Causal Cartographer framework. First, we introduce a graph retrieval-augmented generation agent tasked to retrieve causal relationships from data. This approach allows us to construct a large network of real-world causal relationships that can serve as a repository of causal knowledge and build real-world counterfactuals. In addition, we create a counterfactual reasoning agent constrained by causal relationships to perform reliable step-by-step causal inference. We show that our approach can extract causal knowledge and improve the robustness of LLMs for causal reasoning tasks while reducing inference costs and spurious correlations.

Bias in machine learning models remains a critical challenge, particularly in datasets with numeric features where discrimination may be subtle and hard to detect. Existing fairness frameworks rely on expert knowledge of marginalized groups, such as specific racial groups, and categorical features defining them. Furthermore, most frameworks evaluate bias in models rather than datasets, despite the fact that model bias can often be traced back to dataset shortcomings. Our research aims to remedy this gap by capturing dataset flaws in a set of meta-features at the dataset level, and to warn practitioners of bias risk when using such datasets for model training. We neither restrict the feature type nor expect domain knowledge. To this end, we develop methods to synthesize biased datasets and extend current fairness metrics to continuous features in order to quantify dataset-level discrimination risks. Our approach constructs a meta-database of diverse datasets, from which we derive transferable meta-features that capture dataset properties indicative of bias risk. Our findings demonstrate that dataset-level characteristics can serve as cost-effective indicators of bias risk, providing a novel method for data auditing that does not rely on expert knowledge. This work lays the foundation for early-warning systems, moving beyond model-focused assessments toward a data-centric approach.

In the digital landscape, where fake news proliferates across online platforms, it not only distorts public discourse but also paves the way for sophisticated cyber threats, particularly phishing attacks. These attacks often exploit the credibility crisis created by fake news, tricking individuals with emails that mimic legitimate offers, such as significant discounts, thereby luring them into clicking on malicious links designed to steal personal information. Addressing the nuanced challenge posed by phishing tactics that leverage the deceptive power of fake news, this paper introduces a novel approach: utilizing the Text-to-Text Transfer Transformer (T5) model to improve the phishing detection system. Our research zeroes in on a critical vulnerability exploited by cybercriminals—the use of phishing URLs that are visually and contextually similar to legitimate ones, a tactic made more effective in the disinformation-rich environment fostered by fake news. By employing the T5 model’s advanced text processing capabilities, we aim to both generate and identify phishing URLs that closely mimic legitimate sites, thus enriching anti-phishing databases and boosting the efficacy of detection tools. Our evaluations using the Structural Similarity Index (SSIM) and VirusTotal reveal compelling results: the generated phishing URLs not only achieve an average SSIM score of 90%, indicating a strong visual resemblance to legitimate URLs, but also evade detection effectively, with 96.8% of these URLs marked as "harmless" or "undetected" by VirusTotal, and only a minimal fraction flagged as "malicious". These findings highlight the robustness of our approach in enhancing phishing detection and combating cyber threats in a fake news-infused environment.

Abstract Background Forecasting models predicting trends in hospitalization rates have the potential to inform hospital management during seasonal epidemics of respiratory diseases and the associated surges caused by acute hospital admissions. Hospital bed requirements for elective surgery could be better planned if it were possible to foresee upcoming peaks in severe respiratory illness admissions. Forecasting models can also guide the use of intervention strategies to decrease the spread of respiratory pathogens and thus prevent local health system overload. In this study, we explore the capability of forecasting models to predict the number of hospital admissions in Auckland, New Zealand, within a three-week time horizon. Furthermore, we evaluate probabilistic forecasts and the impact on model performance when integrating laboratory data describing the circulation of respiratory viruses. Methods The dataset used for this exploration results from active hospital surveillance, in which the World Health Organization Severe Acute Respiratory Infection (SARI) case definition was consistently used. This research nurse-led surveillance has been implemented in two public hospitals in Auckland and provides a systematic laboratory testing of SARI patients for nine respiratory viruses, including influenza, respiratory syncytial virus, and rhinovirus. The forecasting strategies used comprise automatic machine learning, one of the most recent generative pre-trained transformers, and established artificial neural network algorithms capable of univariate and multivariate forecasting. Results We found that machine learning models compute more accurate forecasts in comparison to naïve seasonal models. Furthermore, we analyzed the impact of reducing the temporal resolution of forecasts, which decreased the model error of point forecasts and made probabilistic forecasting more reliable. An additional analysis that used the laboratory data revealed strong season-to-season variations in the incidence of respiratory viruses and how this correlates with total hospitalization cases. These variations could explain why it was not possible to improve forecasts by integrating this data. Conclusions Active SARI surveillance and consistent data collection over time enable these data to be used to predict hospital bed utilization. These findings show the potential of machine learning as support for informing systems for proactive hospital management.

Large language models (LLMs) are powerful decision-making tools widely adopted in healthcare, finance, and transportation. Embracing the opportunities and innovations of LLMs is inevitable. However, LLMs inherit stereotypes, misrepresentations, discrimination, and societies' biases from various sources-including training data, algorithm design, and user interactions-resulting in concerns about equality, diversity, and fairness. The bias problem has triggered increased research towards defining, detecting and quantifying bias and developing debiasing techniques. The predominant focus in tackling the bias problem is skewed towards resource-rich regions such as the US and Europe, resulting in a scarcity of research in other societies. As a small country with a unique history, culture and social composition, there is an opportunity for Aotearoa New Zealand's (NZ) research community to address this inadequacy. This paper presents an experimental evaluation of existing bias metrics and debiasing techniques in the NZ context. Research gaps derived from the study and a literature review are outlined, current and ongoing research in this space are discussed, and the suggested scope of research opportunities for NZ are presented.

Despite impressive performance on language modelling and complex reasoning tasks, Large Language Models (LLMs) fall short on the same tasks in uncommon settings or with distribution shifts, exhibiting a lack of generalisation ability. By contrast, systems such as causal models, that learn abstract variables and causal relationships, can demonstrate increased robustness against changes in the distribution. One reason for this success is the existence and use of Independent Causal Mechanisms (ICMs) representing high-level concepts that only sparsely interact. In this work, we apply two concepts from causality to learn ICMs within LLMs. We develop a new LLM architecture composed of multiple sparsely interacting language modelling modules. We show that such causal constraints can improve out-of-distribution performance on abstract and causal reasoning tasks. We also investigate the level of independence and domain specialisation and show that LLMs rely on pre-trained partially domain-invariant mechanisms resilient to fine-tuning.

Large Language Models have shown tremendous performance on a large variety of natural language processing tasks, ranging from text comprehension to common sense reasoning. However, the mechanisms responsible for this success remain opaque, and it is unclear whether LLMs can achieve human-like cognitive capabilities or whether these models are still fundamentally circumscribed. Abstract reasoning is a fundamental task for cognition, consisting of finding and applying a general pattern from few data. Evaluating deep neural architectures on this task could give insight into their potential limitations regarding reasoning and their broad generalisation abilities, yet this is currently an under-explored area. In this paper, we introduce a new benchmark for evaluating language models beyond memorization on abstract reasoning tasks. We perform extensive evaluations of state-of-the-art LLMs, showing that they currently achieve very limited performance in contrast with other natural language tasks, even when applying techniques that have been shown to improve performance on other NLP tasks. We argue that guiding LLM generation to follow causal paths could help improve the generalisation and reasoning abilities of LLMs.

Better understanding the natural world is a crucial task with a wide range of applications. In environments with close proximity between humans and animals, such as zoos, it is essential to better understand the causes behind animal behaviour and what interventions are responsible for changes in their behaviours. This can help to predict unusual behaviours, mitigate detrimental effects and increase the well-being of animals. There has been work on modelling the dynamics behind swarms of birds and insects but the complex social behaviours of mammalian groups remain less explored. In this work, we propose a method to build behavioural models using causal structure discovery and graph neural networks for time series. We apply this method to a mob of meerkats in a zoo environment and study its ability to predict future actions and model the behaviour distribution at an individual-level and at a group level. We show that our method can match and outperform standard deep learning architectures and generate more realistic data, while using fewer parameters and providing increased interpretability.

Causal structure discovery methods are commonly applied to structured data where the causal variables are known and where statistical testing can be used to assess the causal relationships. By contrast, recovering a causal structure from unstructured natural language data such as news articles contains numerous challenges due to the absence of known variables or counterfactual data to estimate the causal links. Large Language Models (LLMs) have shown promising results in this direction but also exhibit limitations. This work investigates LLM's abilities to build causal graphs from text documents and perform counterfactual causal inference. We propose an end-to-end causal structure discovery and causal inference method from natural language: we first use an LLM to extract the instantiated causal variables from text data and build a causal graph. We merge causal graphs from multiple data sources to represent the most exhaustive set of causes possible. We then conduct counterfactual inference on the estimated graph. The causal graph conditioning allows reduction of LLM biases and better represents the causal estimands. We use our method to show that the limitations of LLMs in counterfactual causal reasoning come from prediction errors and propose directions to mitigate them. We demonstrate the applicability of our method on real-world news articles.

Deep neural networks can obtain impressive performance on various tasks under the assumption that their training domain is identical to their target domain. Performance can drop dramatically when this assumption does not hold. One explanation for this discrepancy is the presence of spurious domain-specific correlations in the training data that the network exploits. Causal mechanisms, in the other hand, can be made invariant under distribution changes as they allow disentangling the factors of distribution underlying the data generation. Yet, learning causal mechanisms to improve out-of-distribution generalisation remains an under-explored area. We propose a Bayesian neural architecture that disentangles the learning of the the data distribution from the inference process mechanisms. We show theoretically and experimentally that our model approximates reasoning under causal interventions. We demonstrate the performance of our method, outperforming point estimate-counterparts, on out-of-distribution image recognition tasks where the data distribution acts as strong adversarial confounders.

Deep learning methods often suffer performance degradation due to domain shift, where discrepancies exist between training and testing data distributions. Domain generalization mitigates this problem by leveraging information from multiple source domains to enhance model generalization capabilities for unseen domains. However, existing domain generalization methods typically present examples to the model in a random manner, overlooking the potential benefits of structured data presentation. To bridge this gap, we propose a novel learning strategy, Symmetric Self-Paced Learning (SSPL), for domain generalization. SSPL consists of a Symmetric Self-Paced training scheduler and a Gradient-based Difficulty Measure (GDM). Specifically, the proposed training scheduler initially focuses on easy examples, gradually shifting emphasis to harder examples as training progresses. GDM dynamically evaluates example difficulty through the gradient magnitude with respect to the example itself. Experiments across five popular benchmark datasets demonstrate the effectiveness of the proposed learning strategy.

Growing concern about climate change has intensified efforts to use renewable energy, with wind energy highlighted as a growing source. It is known that wind turbines are characterized by distinct operating modes that reflect production efficiency. In this work, we focus on the forecasting problem for univariate discrete-valued time series of operating modes. We define three prediction strategies to overcome the difficulties associated with missing data. These strategies are evaluated through experiments using five forecasting methods across two real-life datasets. Two of the forecasting methods have been introduced in the statistical literature as extensions of the well-known context algorithm: variable length Markov chains and Bayesian context tree. Additionally, we consider a Bayesian method based on conditional tensor factorization and two different smoothers from the classical tools for time series forecasting. After evaluating each pair prediction strategy/forecasting method in terms of prediction accuracy versus computational complexity, we provide guidance on the methods that are suitable for forecasting the time series of operating modes. The prediction results that we report demonstrate that high accuracy can be achieved with reduced computational resources.

The Diversity, Equity and Inclusion (DEI) initiative started as the Diversity/Inclusion initiative in 2020 [4]. The current report summarizes our activities in 2022. Our responsibility as a community is to ensure that attendees of DB conferences feel included, irrespective of their scientific perspective and personal background. One of the first steps was to establish the role of the DEI chairs at DB Conferences, with the DEI team dedicated to providing leadership to help our community achieve this goal. In this leadership role, the DEI team is advising DEI chairs at DB conferences, serving as a memory of DEI events at conferences, building an agreed-upon vision, and committing to working together to devise a set of measures for achieving DEI. That is pursued via actions led by our core members (Figure 1) and liaisons of individual executive bodies (Figure 2): REACH OUT collects data and experiences from our community. INCLUDE monitors and recommends inclusion efforts. ORGANIZE focuses on in-conference organization efforts, such as adopting a code of conduct. INFORM communicates through various channels. SUPPORT coordinates DEI support from executive bodies and sponsors. SCOUT collates DEI efforts from other communities. COORDINATE manages all actions. Two new actions: MEDIA preserves and disseminates the digital media produced by DEI@DB events. ETHICS establishes and promotes ethics guidelines for publications in our community.

Machine learning models are increasingly used in fields that require high reliability such as cybersecurity. However, these models remain vulnerable to various attacks, among which the adversarial label-flipping attack poses significant threats. In label-flipping attacks, the adversary maliciously flips a portion of training labels to compromise the machine learning model. This paper raises significant concerns as these attacks can camouflage a highly skewed dataset as an easily solvable classification problem, often misleading machine learning practitioners into lower defenses and miscalculations of potential risks. This concern amplifies in tabular data settings, where identifying true labels requires expertise, allowing malicious label-flipping attacks to easily slip under the radar. To demonstrate this risk is inherited in the adversary's objective, we propose FALFA (Fast Adversarial Label-Flipping Attack), a novel efficient attack for crafting adversarial labels. FALFA is based on transforming the adversary's objective and employs linear programming to reduce computational complexity. Using ten real-world tabular datasets, we demonstrate FALFA's superior attack potential, highlighting the need for robust defenses against such threats.

Over 50% of dementia in NZ may be due to high prevalence of risk factors such as hypertension, diabetes and cardiovascular disease. If so, a better understanding of the relationships may present ways to mitigate the risk of dementia. Community-based epidemiological studies would provide definitive evidence but are resource-intensive so we need to explore alternative sources of data such as routinely collected health information. The overarching aim of our research is to use a machine learning model to identify population groups at high risk of cognitive decline and/or dementia. But first we need to gauge the appetite of service users to have their personal data used in this way. This study aimed to explore the attitudes and preferences of people aged 55+ living in Counties Manukau regarding the use of their de-identified health data and understanding their concerns and degree of comfort in different scenarios. An anonymous online survey was conducted with people aged 55+ from NZ European, Māori, Pacific, Chinese and Indian communities currently engaged with older adult health services in Counties Manukau District in 2022. We offered service users the opportunity to complete the survey by telephone if preferred, and the opportunity to be interviewed in languages other than English to ensure adequate representation of the local population. Over 85% of the participants were aware of how their health information was used for service development and assessing population trends, and about 80% were comfortable or very comfortable with data being used in this way, assuming data were stored securely and confidentiality was assured. Participants generally felt that patients’ permission was required if data were being analysed by people other than health professionals or being used by commercial companies, and that there should be transparency about these uses. There were some specific cultural concerns about patients’ data being used after their death. This study supports the idea that the local population are generally supportive of researchers using their de-identified routinely collected data for research purposes that will benefit the population from which the data were taken, but specific cultural concerns need to be addressed.