Search

With the growing adoption of reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs), the risk of backdoor installation during alignment has increased, leading to unintended and harmful behaviors. Existing backdoor triggers are typically limited to fixed word patterns, making them detectable during data cleaning and easily removable post-poisoning. In this work, we explore the use of prompt-specific paraphrases as backdoor triggers, enhancing their stealth and resistance to removal during LLM alignment. We propose AdvBDGen, an adversarially fortified generative fine-tuning framework that automatically generates prompt-specific backdoors that are effective, stealthy, and transferable across models. AdvBDGen employs a generator-discriminator pair, fortified by an adversary, to ensure the installability and stealthiness of backdoors. It enables the crafting and successful installation of complex triggers using as little as 3% of the fine-tuning data. Once installed, these backdoors can jailbreak LLMs during inference, demonstrate improved stability against perturbations compared to traditional constant triggers, and are more challenging to remove. These findings underscore an urgent need for the research community to develop more robust defenses against adversarial backdoor threats in LLM alignment., Comment: Published at the Neurips Safe Generative AI Workshop 2024

Few-shot learning (FSL) aims to enable models to recognize novel objects or classes with limited labelled data. Feature generators, which synthesize new data points to augment limited datasets, have emerged as a promising solution to this challenge. This paper investigates the effectiveness of feature generators in enhancing the embedding process for FSL tasks. To address the issue of inaccurate embeddings due to the scarcity of images per class, we introduce a feature generator that creates visual features from class-level textual descriptions. By training the generator with a combination of classifier loss, discriminator loss, and distance loss between the generated features and true class embeddings, we ensure the generation of accurate same-class features and enhance the overall feature representation. Our results show a significant improvement in accuracy over baseline methods, with our approach outperforming the baseline model by 10% in 1-shot and around 5% in 5-shot approaches. Additionally, both visual-only and visual + textual generators have also been tested in this paper. The code is publicly available at https://github.com/heethanjan/Feature-Generator-for-FSL., Comment: 16 pages, Accepted to ACCV 2024

In this article, we present the bivariate and multivariate functional Moran's I statistics and multivariate functional areal spatial principal component analysis (mfasPCA). These methods are the first of their kind in the field of multivariate areal spatial functional data analysis. The multivariate functional Moran's I statistic is employed to assess spatial autocorrelation, while mfasPCA is utilized for dimension reduction in both univariate and multivariate functional areal data. Through simulation studies and real-world examples, we demonstrate that the multivariate functional Moran's I statistic and mfasPCA are powerful tools for evaluating spatial autocorrelation in univariate and multivariate functional areal data analysis.

Large Language Models (LLMs) have demonstrated impressive capabilities in generating diverse and contextually rich text. However, concerns regarding copyright infringement arise as LLMs may inadvertently produce copyrighted material. In this paper, we first investigate the effectiveness of watermarking LLMs as a deterrent against the generation of copyrighted texts. Through theoretical analysis and empirical evaluation, we demonstrate that incorporating watermarks into LLMs significantly reduces the likelihood of generating copyrighted content, thereby addressing a critical concern in the deployment of LLMs. Additionally, we explore the impact of watermarking on Membership Inference Attacks (MIAs), which aim to discern whether a sample was part of the pretraining dataset and may be used to detect copyright violations. Surprisingly, we find that watermarking adversely affects the success rate of MIAs, complicating the task of detecting copyrighted text in the pretraining dataset. Finally, we propose an adaptive technique to improve the success rate of a recent MIA under watermarking. Our findings underscore the importance of developing adaptive methods to study critical problems in LLMs with potential legal implications., Comment: 21 pages, 6 figures

This study introduces an innovative methodology for mortality forecasting, which integrates signature-based methods within the functional data framework of the Hyndman-Ullah (HU) model. This new approach, termed the Hyndman-Ullah with truncated signatures (HUts) model, aims to enhance the accuracy and robustness of mortality predictions. By utilizing signature regression, the HUts model aims to capture complex, nonlinear dependencies in mortality data which enhances forecasting accuracy across various demographic conditions. The model is applied to mortality data from 12 countries, comparing its forecasting performance against classical models like the Lee-Carter model and variants of the HU models across multiple forecast horizons. Our findings indicate that overall the HUts model not only provides more precise point forecasts but also shows robustness against data irregularities, such as those observed in countries with historical outliers. The integration of signature-based methods enables the HUts model to capture complex patterns in mortality data, making it a powerful tool for actuaries and demographers. Prediction intervals are also constructed using bootstrapping methods., Comment: 46 pages, 24 figures, 9 tables

In this paper, we explore dimension reduction for time series of functional data within both stationary and non-stationary frameworks. We introduce a functional framework of generalized dynamic principal component analysis (GDPCA). The concept of GDPCA aims for better adaptation to possible nonstationary features of the series. We define the functional generalized dynamic principal component (GDPC) as static factor time series in a functional dynamic factor model and obtain the multivariate GDPC from a truncation of the functional dynamic factor model. GDFPCA uses a minimum squared error criterion to evaluate the reconstruction of the original functional time series. The computation of GDPC involves a two-step estimation of the coefficient vector of the loading curves in a basis expansion. We provide a proof of the consistency of the reconstruction of the original functional time series with GDPC converging in mean square to the original functional time series. Monte Carlo simulation studies indicate that the proposed GDFPCA is comparable to dynamic functional principal component analysis (DFPCA) when the data generating process is stationary, and outperforms DFPCA and FPCA when the data generating process is non-stationary. The results of applications to real data reaffirm the findings in simulation studies.

Recent advancements in Reinforcement Learning with Human Feedback (RLHF) have significantly impacted the alignment of Large Language Models (LLMs). The sensitivity of reinforcement learning algorithms such as Proximal Policy Optimization (PPO) has led to new line work on Direct Policy Optimization (DPO), which treats RLHF in a supervised learning framework. The increased practical use of these RLHF methods warrants an analysis of their vulnerabilities. In this work, we investigate the vulnerabilities of DPO to poisoning attacks under different scenarios and compare the effectiveness of preference poisoning, a first of its kind. We comprehensively analyze DPO's vulnerabilities under different types of attacks, i.e., backdoor and non-backdoor attacks, and different poisoning methods across a wide array of language models, i.e., LLama 7B, Mistral 7B, and Gemma 7B. We find that unlike PPO-based methods, which, when it comes to backdoor attacks, require at least 4\% of the data to be poisoned to elicit harmful behavior, we exploit the true vulnerabilities of DPO more simply so we can poison the model with only as much as 0.5\% of the data. We further investigate the potential reasons behind the vulnerability and how well this vulnerability translates into backdoor vs non-backdoor attacks.

Image thumbnails are a valuable data source for fixation filtering, scanpath classification, and visualization of eye tracking data. They are typically extracted with a constant size, approximating the foveated area. As a consequence, the focused area of interest in the scene becomes less prominent in the thumbnail with increasing distance, affecting image-based analysis techniques. In this work, we propose depth-adaptive thumbnails, a method for varying image size according to the eye-to-object distance. Adjusting the visual angle relative to the distance leads to a zoom effect on the focused area. We evaluate our approach on recordings in augmented reality, investigating the similarity of thumbnails and scanpaths. Our quantitative findings suggest that considering the eye-to-object distance improves the quality of data analysis and visualization. We demonstrate the utility of depth-adaptive thumbnails for applications in scanpath comparison and visualization.

In this work, we investigate the means of using curiosity on replay buffers to improve offline multi-task continual reinforcement learning when tasks, which are defined by the non-stationarity in the environment, are non labeled and not evenly exposed to the learner in time. In particular, we investigate the use of curiosity both as a tool for task boundary detection and as a priority metric when it comes to retaining old transition tuples, which we respectively use to propose two different buffers. Firstly, we propose a Hybrid Reservoir Buffer with Task Separation (HRBTS), where curiosity is used to detect task boundaries that are not known due to the task agnostic nature of the problem. Secondly, by using curiosity as a priority metric when it comes to retaining old transition tuples, a Hybrid Curious Buffer (HCB) is proposed. We ultimately show that these buffers, in conjunction with regular reinforcement learning algorithms, can be used to alleviate the catastrophic forgetting issue suffered by the state of the art on replay buffers when the agent's exposure to tasks is not equal along time. We evaluate catastrophic forgetting and the efficiency of our proposed buffers against the latest works such as the Hybrid Reservoir Buffer (HRB) and the Multi-Time Scale Replay Buffer (MTR) in three different continual reinforcement learning settings. Experiments were done on classical control tasks and Metaworld environment. Experiments show that our proposed replay buffers display better immunity to catastrophic forgetting compared to existing works in most of the settings.

Stock market indices are volatile by nature, and sudden shocks are known to affect volatility patterns. The autoregressive conditional heteroskedasticity (ARCH) and generalized ARCH (GARCH) models neglect structural breaks triggered by sudden shocks that may lead to an overestimation of persistence, causing an upward bias in the estimates. Different regime-switching models that have abrupt regime-switching governed by a Markov chain were developed to model volatility in financial time series data. Volatility modelling was also extended to spatially interconnected time series, resulting in spatial variants of ARCH models. This inspired us to propose a Markov switching framework of the spatio-temporal log-ARCH model. In this article, we discuss the Markov-switching extension of the model, the estimation procedure and the smooth inferences of the regimes. The Monte-Carlo simulation studies show that the maximum likelihood estimation method for our proposed model has good finite sample properties. The proposed model was applied to 28 stock indices data that were presumably affected by the 2015-2016 Chinese stock market crash. The results showed that our model is a better fit compared to that of the one-regime counterpart. Furthermore, the smoothed inference of the data indicated the approximate periods where structural breaks occurred. This model can capture structural breaks that simultaneously occur in nearby locations.

This work proposes a functional data analysis approach for morphometrics with applications in classifying three shrew species (S. murinus, C. monticola and C. malayana) based on the images. The discrete landmark data of craniodental views (dorsal, jaw and lateral) are converted into continuous curves where the curves are represented as linear combinations of basis functions. A comparative study based on four machine learning algorithms such as naive Bayes, support vector machine, random forest, and generalized linear models was conducted on the predicted principal component scores obtained from the FDA approach and classical approach (combination of all three craniodental views and individual views). The FDA approach produced better results in separating the three clusters of shrew species compared to the classical method and the dorsal view gave the best representation in classifying the three shrew species. Overall, based on the FDA approach, GLM of the predicted PCA scores was the most accurate (95.4% accuracy) among the four classification models., Comment: 24 pages,4 Figures,2 tables,Dharini Pathmanathan (corresponding author)

Augmented reality (AR) technology provides means for embedding visualization in a real-world context. Such techniques allow situated analyses of live data in their spatial domain. However, as existing techniques have to be adapted for this context and new approaches will be developed, the evaluation thereof poses new challenges for researchers. Apart from established performance measures, eye tracking has proven to be a valuable means to assess visualizations qualitatively and quantitatively. We discuss the challenges and opportunities of eye tracking for the evaluation of situated visualizations. We envision that an extension of gaze-based evaluation methodology into this field will provide new insights on how people perceive and interact with visualizations in augmented reality.

This study analyzes and compares the drivetrain losses of two methods of fuel cell integration in electric vehicle drivetrains. The first is a conventional (boosted) two-stage system while the second is a dual inverter based solution. Each source of drivetrain losses is described by mathematical equations and the impact of higher order harmonics are observed through circuit model simulation. The dual inverter system achieves an overall energy efficiency improvement of 5.27% and 10.13% for highway and urban driving compared to the conventional method. The use of lower voltage rated power modules and lower switching frequency in the dual inverter system has significantly reduced the switching losses and improved the driving efficiency.

Objective: Continuous EEG (cEEG) monitoring is associated with lower mortality in critically ill patients, however it is underutilized due to the difficulty of manually interpreting prolonged streams of cEEG data. Here we present a novel real-time, machine learning-based alerting and monitoring system for epilepsy and seizures (RAMSES) that dramatically reduces the amount of manual EEG review. Methods: We developed a custom data reduction algorithm using a random forest, and deployed it within an online cloud-based platform which streams data and communicates interactively with caregivers via a web interface to display algorithm results. We validate RAMSES on cEEG recordings from 77 patients undergoing routine scalp ICU EEG monitoring. Results: On subjects with seizures we achieved >80% overall data reduction, while detecting a mean of 84% of seizures across all validation patients, with 19/27 patients achieving 100% seizure detection. On seizure free-patients, the majority of cEEG records, we reduced data requiring manual review by >83%. Conclusion: This study validates a platform for machine-learning assisted data reduction. Significance: This work represents a meaningful step towards improving utility and decreasing cost for cEEG monitoring We also make our high-quality annotated dataset of 77 ICU cEEG recordings public for others to validate and improve upon our methods.

This paper considers a multi-armed bandit (MAB) problem in which multiple mobile agents receive rewards by sampling from a collection of spatially dispersed stochastic processes, called bandits. The goal is to formulate a decentralized policy for each agent, in order to maximize the total cumulative reward over all agents, subject to option availability and inter-agent communication constraints. The problem formulation is motivated by applications in which a team of autonomous mobile robots cooperates to accomplish an exploration and exploitation task in an uncertain environment. Bandit locations are represented by vertices of the spatial graph. At any time, an agent's option consist of sampling the bandit at its current location, or traveling along an edge of the spatial graph to a new bandit location. Communication constraints are described by a directed, non-stationary, stochastic communication graph. At any time, agents may receive data only from their communication graph in-neighbors. For the case of a single agent on a fully connected spatial graph, it is known that the expected regret for any optimal policy is necessarily bounded below by a function that grows as the logarithm of time. A class of policies called upper confidence bound (UCB) algorithms asymptotically achieve logarithmic regret for the classical MAB problem. In this paper, we propose a UCB-based decentralized motion and option selection policy and a non-stationary stochastic communication protocol that guarantee logarithmic regret. To our knowledge, this is the first such decentralized policy for non-fully connected spatial graphs with communication constraints. When the spatial graph is fully connected and the communication graph is stationary, our decentralized algorithm matches or exceeds the best reported prior results from the literature., Comment: Pre-print submitted for review to the 2020 CDC

Uncertainty quantification (UQ) is a vital step in using mathematical models and simulations to take decisions. The field of cardiac simulation has begun to explore and adopt UQ methods to characterise uncertainty in model inputs and how that propagates through to outputs or predictions. In this perspective piece we draw attention to an important and under-addressed source of uncertainty in our predictions -- that of uncertainty in the model structure or the equations themselves. The difference between imperfect models and reality is termed model discrepancy, and we are often uncertain as to the size and consequences of this discrepancy. Here we provide two examples of the consequences of discrepancy when calibrating models at the ion channel and action potential scales. Furthermore, we attempt to account for this discrepancy when calibrating and validating an ion channel model using different methods, based on modelling the discrepancy using Gaussian processes (GPs) and autoregressive-moving-average (ARMA) models, then highlight the advantages and shortcomings of each approach. Finally, suggestions and lines of enquiry for future work are provided., Comment: This version is published in Philosophical Transactions of the Royal Society A; Updated in response to reviewer comments, including: added details to the introduction, fixed mathematical notations for clarity, and moved the original Table 3 to the supplement to avoid confusion

Sleep apnea is a breathing disorder where a person repeatedly stops breathing in sleep. Early detection is crucial for infants because it might bring long term adversities. The existing accurate detection mechanism (pulse oximetry) is a skin contact measurement. The existing non-contact mechanisms (acoustics, video processing) are not accurate enough. This paper presents a novel algorithm for the detection of sleep apnea with video processing. The solution is non-contact, accurate and lightweight enough to run on a single board computer. The paper discusses the accuracy of the algorithm on real data, advantages of the new algorithm, its limitations and suggests future improvements., Comment: Gihan Jayatilaka, Harshana Weligampola and Suren Sritharan are equally contributing authors

This paper proposes a novel policy for a group of agents to, individually as well as collectively, solve a multi armed bandit (MAB) problem. The policy relies solely on the information that an agent has obtained through sampling of the options on its own and through communication with neighbors. The option selection policy is based on an Upper Confidence Based (UCB) strategy while the communication strategy that is proposed forces agents to communicate with other agents who they believe are most likely to be exploring than exploiting. The overall strategy is shown to significantly outperform an independent Erd\H{o}s-R\'{e}nyi (ER) graph based random communication policy. The policy is shown to be cost effective in terms of communication and thus to be easily scalable to a large network of agents., Comment: This is the full version of a preprint that will appear in the proceedings of the 2020 European Control Conference (ECC)

When modelling tissue-level cardiac electrophysiology, continuum approximations to the discrete cell-level equations are used to maintain computational tractability. One of the most commonly used models is represented by the bidomain equations, the derivation of which relies on a homogenisation technique to construct a suitable approximation to the discrete model. This derivation does not explicitly account for the presence of gap junctions connecting one cell to another. It has been seen experimentally [Rohr, Cardiovasc. Res. 2004] that these gap junctions have a marked effect on the propagation of the action potential, specifically as the upstroke of the wave passes through the gap junction. In this paper we explicitly include gap junctions in a both a 2D discrete model of cardiac electrophysiology, and the corresponding continuum model, on a simplified cell geometry. Using these models we compare the results of simulations using both continuum and discrete systems. We see that the form of the action potential as it passes through gap junctions cannot be replicated using a continuum model, and that the underlying propagation speed of the action potential ceases to match up between models when gap junctions are introduced. In addition, the results of the discrete simulations match the characteristics of those shown in Rohr 2004. From this, we suggest that a hybrid model -- a discrete system following the upstroke of the action potential, and a continuum system elsewhere -- may give a more accurate description of cardiac electrophysiology., Comment: In Proceedings HSB 2012, arXiv:1208.3151