Cases of laryngeal cancer are predicted to rise significantly in the coming years. Current diagnostic pathways cause many patients to be incorrectly referred to urgent suspected cancer pathways, putting undue stress on both patients and the medical system. Artificial intelligence offers a promising solution by enabling non-invasive detection of laryngeal cancer from patient speech, which could help prioritise referrals more effectively and reduce inappropriate referrals of non-cancer patients. To realise this potential, open science is crucial. A major barrier in this field is the lack of open-source datasets and reproducible benchmarks, forcing researchers to start from scratch. Our work addresses this challenge by introducing a benchmark suite comprising 36 models trained and evaluated on open-source datasets. These models are accessible in a public repository, providing a foundation for future research. They evaluate three different algorithms and three audio feature sets, offering a comprehensive benchmarking framework. We propose standardised metrics and evaluation methodologies to ensure consistent and comparable results across future studies. The presented models include both audio-only inputs and multimodal inputs that incorporate demographic and symptom data, enabling their application to datasets with diverse patient information. By providing these benchmarks, future researchers can evaluate their datasets, refine the models, and use them as a foundation for more advanced approaches. This work aims to provide a baseline for establishing reproducible benchmarks, enabling researchers to compare new methods against these standards and ultimately advancing the development of AI tools for detecting laryngeal cancer.

Large Language Models (LLMs) have shown impressive performance across a wide range of tasks. However, the size of LLMs is steadily increasing, hindering their application on computationally constrained environments. On the other hand, despite their general capabilities, there are many situations where only one specific task is performed, rendering all other capabilities unnecessary and wasteful. This leads us to the following question: Is it possible to extract the minimal subset from an LLM that is able to perform a specific task in a faster, standalone manner? Recent works on Mechanistic Interpretability (MI) have shown that specific tasks are performed by a localized subset of components, or circuit. However, current techniques used to identify the circuit cannot be used to extract it for its standalone usage. In this work, we propose a novel approach to automatically extract the subset of the LLM that properly performs a targeted task requiring no additional training and a small amount of data samples. We evaluate our approach on different tasks and show that the resulting models are (i) considerably smaller, reducing the number of parameters up to 82.77% and (ii) more interpretable, as they focus on the circuit that is used to carry out the specific task, and can therefore be understood using MI techniques.

Large Automatic Speech Recognition (ASR) models demand a vast number of parameters, copious amounts of data, and significant computational resources during the training process. However, such models can merely be deployed on high-compute cloud platforms and are only capable of performing speech recognition tasks. This leads to high costs and restricted capabilities. In this report, we initially propose the elastic mixture of the expert (eMoE) model. This model can be trained just once and then be elastically scaled in accordance with deployment requirements. Secondly, we devise an unsupervised data creation and validation procedure and gather millions of hours of audio data from diverse domains for training. Using these two techniques, our system achieves elastic deployment capabilities while reducing the Character Error Rate (CER) on the SpeechIO testsets from 4.98\% to 2.45\%. Thirdly, our model is not only competent in Mandarin speech recognition but also proficient in multilingual, multi-dialect, emotion, gender, and sound event perception. We refer to this as Automatic Speech Perception (ASP), and the perception results are presented in the experimental section.

The growing threats of uncertainties, anomalies, and cyberattacks on power grids are driving a critical need to advance situational awareness which allows system operators to form a complete and accurate picture of the present and future state. Simulation and estimation are foundational tools in this process. However, existing tools lack the robustness and efficiency required to achieve the level of situational awareness needed for the ever-evolving threat landscape. Industry-standard (steady-state) simulators are not robust to blackouts, often leading to non-converging or non-actionable results. Estimation tools lack robustness to anomalous data, returning erroneous system states. Efficiency is the other major concern as nonlinearities and scalability issues make large systems slow to converge. This thesis addresses robustness and efficiency gaps through a dual-fold contribution. We first address the inherent limitations in the existing physics-based and data-driven worlds; and then transcend the boundaries of conventional algorithmic design in the direction of a new paradigm -- Physics-ML Synergy -- which integrates the strengths of the two worlds. Our approaches are built on circuit formulation which provides a unified framework that applies to both transmission and distribution. Sparse optimization acts as the key enabler to make these tools intrinsically robust and immune to random threats, pinpointing dominant sources of (random) blackouts and data errors. Further, we explore sparsity-exploiting optimizations to develop lightweight ML models whose prediction and detection capabilities are a complement to physics-based tools; and whose lightweight designs advance generalization and scalability. Finally, Physics-ML Synergy brings robustness and efficiency further against targeted cyberthreats, by interconnecting our physics-based tools with lightweight ML.

Constraint Acquisition (CA) aims to widen the use of constraint programming by assisting users in the modeling process. However, most CA methods suffer from a significant drawback: they learn a single set of individual constraints for a specific problem instance, but cannot generalize these constraints to the parameterized constraint specifications of the problem. In this paper, we address this limitation by proposing GenCon, a novel approach to learn parameterized constraint models capable of modeling varying instances of the same problem. To achieve this generalization, we make use of statistical learning techniques at the level of individual constraints. Specifically, we propose to train a classifier to predict, for any possible constraint and parameterization, whether the constraint belongs to the problem. We then show how, for some classes of classifiers, we can extract decision rules to construct interpretable constraint specifications. This enables the generation of ground constraints for any parameter instantiation. Additionally, we present a generate-and-test approach that can be used with any classifier, to generate the ground constraints on the fly. Our empirical results demonstrate that our approach achieves high accuracy and is robust to noise in the input instances.

Educational stakeholders are often particularly interested in sparse, delayed student outcomes, like end-of-year statewide exams. The rare occurrence of such assessments makes it harder to identify students likely to fail such assessments, as well as making it slow for researchers and educators to be able to assess the effectiveness of particular educational tools. Prior work has primarily focused on using logs from students full usage (e.g. year-long) of an educational product to predict outcomes, or considered predictive accuracy using a few minutes to predict outcomes after a short (e.g. 1 hour) session. In contrast, we investigate machine learning predictors using students' logs during their first few hours of usage can provide useful predictive insight into those students' end-of-school year external assessment. We do this on three diverse datasets: from students in Uganda using a literacy game product, and from students in the US using two mathematics intelligent tutoring systems. We consider various measures of the accuracy of the resulting predictors, including its ability to identify students at different parts along the assessment performance distribution. Our findings suggest that short-term log usage data, from 2-5 hours, can be used to provide valuable signal about students' long-term external performance.

In pseudonymous online fora like Reddit, the benefits of self-disclosure are often apparent to users (e.g., I can vent about my in-laws to understanding strangers), but the privacy risks are more abstract (e.g., will my partner be able to tell that this is me?). Prior work has sought to develop natural language processing (NLP) tools that help users identify potentially risky self-disclosures in their text, but none have been designed for or evaluated with the users they hope to protect. Absent this assessment, these tools will be limited by the social-technical gap: users need assistive tools that help them make informed decisions, not paternalistic tools that tell them to avoid self-disclosure altogether. To bridge this gap, we conducted a study with N = 21 Reddit users; we had them use a state-of-the-art NLP disclosure detection model on two of their authored posts and asked them questions to understand if and how the model helped, where it fell short, and how it could be improved to help them make more informed decisions. Despite its imperfections, users responded positively to the model and highlighted its use as a tool that can help them catch mistakes, inform them of risks they were unaware of, and encourage self-reflection. However, our work also shows how, to be useful and usable, AI for supporting privacy decision-making must account for posting context, disclosure norms, and users' lived threat models, and provide explanations that help contextualize detected risks.

In healthcare, the integration of multimodal data is pivotal for developing comprehensive diagnostic and predictive models. However, managing missing data remains a significant challenge in real-world applications. We introduce MARIA (Multimodal Attention Resilient to Incomplete datA), a novel transformer-based deep learning model designed to address these challenges through an intermediate fusion strategy. Unlike conventional approaches that depend on imputation, MARIA utilizes a masked self-attention mechanism, which processes only the available data without generating synthetic values. This approach enables it to effectively handle incomplete datasets, enhancing robustness and minimizing biases introduced by imputation methods. We evaluated MARIA against 10 state-of-the-art machine learning and deep learning models across 8 diagnostic and prognostic tasks. The results demonstrate that MARIA outperforms existing methods in terms of performance and resilience to varying levels of data incompleteness, underscoring its potential for critical healthcare applications.

High-quality professional development for teachers can facilitate the learning of best teaching practices, which in turn can lead to higher levels of student performance (Borko et al., 2010; Didion et al., 2020; Gore et al., 2021; Hubers et al., 2022; Schelling & Rubenstein, 2023). For instance, feedback on actual practices has proven effective in enhancing teaching methods and subsequently improving student outcomes (Allen et al., 2011; Gregory et al., 2017), even among students not directly taught by the teachers receiving the feedback (Opper, 2019). Thus, focusing on the evaluation of teaching practices to facilitate professional development is essential, as it can lead to improved teaching methods and ultimately to higher levels of student outcomes. Despite its acknowledged importance and the pressures from high-stakes accountability systems, most professional development opportunities remain fragmented and insufficient to meet teachers' needs (Borko, 2004; Hsu & Malkin, 2013). The reason for this may be that, although it is known that teaching practices such as cognitive activation, supportive climate, and classroom management, are relevant for enhancing teaching quality and student outcomes (Xie & Derakhshan, 2021), these dimensions may be too abstract or general, which can hinder the implementation of concrete actions to improve teaching quality. In this regard, evidence suggests that targeting more specific factors for intervention, rather than abstract ones, allows teachers to better understand and change their practices (Soderberg et al., 2015).

This paper considers the problem of fair probabilistic binary classification with binary protected groups. The classifier assigns scores, and a practitioner predicts labels using a certain cut-off threshold based on the desired trade-off between false positives vs. false negatives. It derives these thresholds from the ROC of the classifier. The resultant classifier may be unfair to one of the two protected groups in the dataset. It is desirable that no matter what threshold the practitioner uses, the classifier should be fair to both the protected groups; that is, the $\mathcal{L}_p$ norm between FPRs and TPRs of both the protected groups should be at most $\varepsilon$. We call such fairness on ROCs of both the protected attributes $\varepsilon_p$-Equalized ROC. Given a classifier not satisfying $\varepsilon_1$-Equalized ROC, we aim to design a post-processing method to transform the given (potentially unfair) classifier's output (score) to a suitable randomized yet fair classifier. That is, the resultant classifier must satisfy $\varepsilon_1$-Equalized ROC. First, we introduce a threshold query model on the ROC curves for each protected group. The resulting classifier is bound to face a reduction in AUC. With the proposed query model, we provide a rigorous theoretical analysis of the minimal AUC loss to achieve $\varepsilon_1$-Equalized ROC. To achieve this, we design a linear time algorithm, namely \texttt{FROC}, to transform a given classifier's output to a probabilistic classifier that satisfies $\varepsilon_1$-Equalized ROC. We prove that under certain theoretical conditions, \texttt{FROC}\ achieves the theoretical optimal guarantees. We also study the performance of our \texttt{FROC}\ on multiple real-world datasets with many trained classifiers.