Foundation models such as large language models (LLMs) have shown remarkable capabilities redefining the field of artificial intelligence. At the same time, they present pressing and challenging socio-technical risks regarding the trustworthiness of their outputs and their alignment with human values and ethics [Bommasani et al., 2021]. Evaluating LLMs is therefore a multi-dimensional problem, where those risks are assessed across diverse tasks and domains [Chang et al., 2023]. In order to quantify these risks, Liang et al. [2022], Wang et al. [2023], Huang et al. [2023] proposed benchmarks of automatic metrics for probing the trustworthiness of LLMs. These metrics include accuracy, robustness, fairness, toxicity of the outputs, etc. Human evaluation benchmarks can be even more nuanced, and are often employed when tasks surpass the scope of standard metrics. Notable benchmarks based on human and automatic evaluations include, among others, Chatbot Arena [Zheng et al., 2023], HELM [Bommasani et al., 2023], MosaicML's Eval, Open LLM Leaderboard [Wolf, 2023], and BIG-bench [Srivastava et al., 2022], each catering to specific evaluation areas such as chatbot performance, knowledge assessment, and domain-specific challenges. Traditional metrics, however, sometimes do not correlate well with human judgments.

This work presents a novel reconfigurable architecture for Low Latency Graph Neural Network (LL-GNN) designs for particle detectors, delivering unprecedented low latency performance. Incorporating FPGA-based GNNs into particle detectors presents a unique challenge since it requires sub-microsecond latency to deploy the networks for online event selection with a data rate of hundreds of terabytes per second in the Level-1 triggers at the CERN Large Hadron Collider experiments. This paper proposes a novel outer-product based matrix multiplication approach, which is enhanced by exploiting the structured adjacency matrix and a column-major data layout. Moreover, a fusion step is introduced to further reduce the end-to-end design latency by eliminating unnecessary boundaries. Furthermore, a GNN-specific algorithm-hardware co-design approach is presented which not only finds a design with a much better latency but also finds a high accuracy design under given latency constraints. To facilitate this, a customizable template for this low latency GNN hardware architecture has been designed and open-sourced, which enables the generation of low-latency FPGA designs with efficient resource utilization using a high-level synthesis tool. Evaluation results show that our FPGA implementation is up to 9.0 times faster and achieves up to 13.1 times higher power efficiency than a GPU implementation. Compared to the previous FPGA implementations, this work achieves 6.51 to 16.7 times lower latency. Moreover, the latency of our FPGA design is sufficiently low to enable deployment of GNNs in a sub-microsecond, real-time collider trigger system, enabling it to benefit from improved accuracy. The proposed LL-GNN design advances the next generation of trigger systems by enabling sophisticated algorithms to process experimental data efficiently.

Machine learning models, in particular deep neural networks, are currently an integral part of various applications, from healthcare to finance. However, using sensitive data to train these models raises concerns about privacy and security. One method that has emerged to verify if the trained models are privacy-preserving is Membership Inference Attacks (MIA), which allows adversaries to determine whether a specific data point was part of a model's training dataset. While a series of MIAs have been proposed in the literature, only a few can achieve high True Positive Rates (TPR) in the low False Positive Rate (FPR) region (0.01%~1%). This is a crucial factor to consider for an MIA to be practically useful in real-world settings. In this paper, we present a novel approach to MIA that is aimed at significantly improving TPR at low FPRs. Our method, named learning-based difficulty calibration for MIA(LDC-MIA), characterizes data records by their hardness levels using a neural network classifier to determine membership. The experiment results show that LDC-MIA can improve TPR at low FPR by up to 4x compared to the other difficulty calibration based MIAs. It also has the highest Area Under ROC curve (AUC) across all datasets. Our method's cost is comparable with most of the existing MIAs, but is orders of magnitude more efficient than one of the state-of-the-art methods, LiRA, while achieving similar performance.

Active learning is a commonly used approach that reduces the labeling effort required to train deep neural networks. However, the effectiveness of current active learning methods is limited by their closed-world assumptions, which assume that all data in the unlabeled pool comes from a set of predefined known classes. This assumption is often not valid in practical situations, as there may be unknown classes in the unlabeled data, leading to the active open-set annotation problem. The presence of unknown classes in the data can significantly impact the performance of existing active learning methods due to the uncertainty they introduce. To address this issue, we propose a novel data-centric active learning method called NEAT that actively annotates open-set data. NEAT is designed to label known classes data from a pool of both known and unknown classes unlabeled data. It utilizes the clusterability of labels to identify the known classes from the unlabeled pool and selects informative samples from those classes based on a consistency criterion that measures inconsistencies between model predictions and local feature distribution. Unlike the recently proposed learning-centric method for the same problem, NEAT is much more computationally efficient and is a data-centric active open-set annotation method. Our experiments demonstrate that NEAT achieves significantly better performance than state-of-the-art active learning methods for active open-set annotation.

The development of cancer is triggered by alterations and mutations in the DNA. The majority of DNA changes responsible for cancer occur within specific regions known as genes. Among the various types of cancers, skin cancer is among the five on the list. If we disregard breast and prostate cancer which are gender-dependent, skin cancer will remain in the third largest cancer category among many others. Based on the statistics released by the American Cancer Society (ACS) [1], there were 58,120 recorded cases of skin cancer among males and 39,490 cases among females. An intriguing observation is that the incidence of skin cancer has been steadily rising from 1992 to 2019, with a notable exception in 2020 [2]. This exception can be attributed to the understandable decrease in cases during the COVID-19 pandemic, as people were mostly confined to their homes. This decline is reasonable considering that exposure to ultraviolet (UV) radiation is a significant contributing factor to the development of skin cancer. More people are diagnosed with skin cancer each year in the U.S. than all other cancers combined [3].

Bayesian Neural Networks (BayNNs) can inherently estimate predictive uncertainty, facilitating informed decision-making. Dropout-based BayNNs are increasingly implemented in spintronics-based computation-in-memory architectures for resourceconstrained yet high-performance safety-critical applications. Although uncertainty estimation is important, the reliability of Dropout generation and BayNN computation is equally important for target applications but is overlooked in existing works. However, testing BayNNs is significantly more challenging compared to conventional NNs, due to their stochastic nature. In this paper, we present for the first time the model of the non-idealities of the spintronics-based Dropout module and analyze their impact on uncertainty estimates and accuracy. Furthermore, we propose a testing framework based on repeatability ranking for Dropout-based BayNN with up to 100% fault coverage while using only 0.2% of training data as test vectors. Bayesian Neural Networks (BayNNs) offer substantial benefits over conventional neural networks (NNs), particularly in safety-critical applications where reliability and confidence in prediction are paramount [1]. Unlike traditional NNs, BayNNs can inherently capture and estimate the uncertainty of their predictions, enhancing decision-making under uncertain conditions. However, their implementation faces significant computational bottlenecks, especially on edge devices. Spintronics-based computation-in-memory (Spintronics-CIM) architectures are a promising solution for the hardware realization of BayNNs as they mitigate some of the inherent computational costs, balancing high-performance demands with the constraints of resourcelimited devices.

Despite recent advances, Automatic Speech Recognition (ASR) systems are still far from perfect. Typical errors include acronyms, named entities and domain-specific special words for which little or no data is available. To address the problem of recognizing these words, we propose an self-supervised continual learning approach. Given the audio of a lecture talk with corresponding slides, we bias the model towards decoding new words from the slides by using a memory-enhanced ASR model from previous work. Then, we perform inference on the talk, collecting utterances that contain detected new words into an adaptation dataset. Continual learning is then performed on this set by adapting low-rank matrix weights added to each weight matrix of the model. The whole procedure is iterated for many talks. We show that with this approach, we obtain increasing performance on the new words when they occur more frequently (more than 80% recall) while preserving the general performance of the model.

While extensive guidance exists for ensuring the reproducibility of one's own study, there is little discussion regarding the reproduction and replication of external studies within one's own research. To initiate this discussion, drawing lessons from our experience reproducing an operational product for predicting tropical cyclogenesis, we present a two-dimensional framework to offer guidance on reproduction and replication. Our framework, representing model fitting on one axis and its use in inference on the other, builds upon three key aspects: the dataset, the metrics, and the model itself. By assessing the trajectories of our studies on this 2D plane, we can better inform the claims made using our research. Additionally, we use this framework to contextualize the utility of benchmark datasets in the atmospheric sciences. Our two-dimensional framework provides a tool for researchers, especially early career researchers, to incorporate prior work in their own research and to inform the claims they can make in this context.

Utilizing state-of-the-art Large Language Models (LLMs), automatic code generation models play a pivotal role in enhancing the productivity of software development procedures. As the adoption of LLMs becomes more widespread in software coding ecosystems, a pressing issue has emerged: does the generated code contain social bias and unfairness, such as those related to age, gender, and race? This issue concerns the integrity, fairness, and ethical foundation of software applications that depend on the code generated by these models, yet is under-explored in the literature. This paper presents a novel bias testing framework that is specifically designed for code generation tasks. Based on this framework, we conduct an extensive evaluation of the bias in code generated by five state-of-the-art LLMs. Our findings reveal that 20.29% to 44.93% code functions generated by the models under study are biased when handling bias sensitive tasks (i.e., tasks that involve sensitive attributes such as age and gender). This indicates that the existing LLMs can be unfair in code generation, posing risks of unintended and harmful software behaviors. To mitigate bias for code generation models, we evaluate five bias mitigation prompt strategies, i.e., utilizing bias testing results to refine the code (zero-shot), one-, few-shot, and two Chain-of-Thought (CoT) prompts. Our evaluation results illustrate that these strategies are all effective in mitigating bias. Overall, one-shot and few-shot learning are the two most effective. For GPT-4, 80% to 90% code bias can be removed with one-shot learning.

The goal of Causal Discovery is to find automated search methods for learning causal structures from observational data. In some cases all variables of the interested causal mechanism are measured, and the task is to predict the effects one measured variable has on another. In contrast, sometimes the variables of primary interest are not directly observable but instead inferred from their manifestations in the data. These are referred to as latent variables. One commonly known example is the psychological construct of intelligence, which cannot directly measured so researchers try to assess through various indicators such as IQ tests. In this case, casual discovery algorithms can uncover underlying patterns and structures to reveal the causal connections between the latent variables and between the latent and observed variables. This thesis focuses on two questions in causal discovery: providing an alternative definition of k-Triangle Faithfulness that (i) is weaker than strong faithfulness when applied to the Gaussian family of distributions, (ii) can be applied to non-Gaussian families of distributions, and (iii) under the assumption that the modified version of Strong Faithfulness holds, can be used to show the uniform consistency of a modified causal discovery algorithm; relaxing the sufficiency assumption to learn causal structures with latent variables. Given the importance of inferring cause-and-effect relationships for understanding and forecasting complex systems, the work in this thesis of relaxing various simplification assumptions is expected to extend the causal discovery method to be applicable in a wider range with diversified causal mechanism and statistical phenomena.