The growing demand for accurate and equitable AI models in digital dermatology faces a significant challenge: the lack of diverse, high-quality labeled data. In this work, we investigate the potential of domain-specific foundation models for dermatology in addressing this challenge. We utilize self-supervised learning (SSL) techniques to pre-train models on a dataset of over 240,000 dermatological images from public and private collections. Our study considers several SSL methods and compares the resulting foundation models against domain-agnostic models like those pre-trained on ImageNet and state-of-the-art models such as MONET across 12 downstream tasks. Unlike previous research, we emphasize the development of smaller models that are more suitable for resource-limited clinical settings, facilitating easier adaptation to a broad range of use cases. Results show that models pre-trained in this work not only outperform general-purpose models but also approach the performance of models 50 times larger on clinically relevant diagnostic tasks. To promote further research in this direction, we publicly release both the training code and the foundation models, which can benefit clinicians in dermatological applications.

Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity per group. In this paper, we study the cause of this inconsistency by unifying existing methods into a standard optimization framework. We show that all methods set proportions to minimize total loss, subject to a method-specific mixing law -- an assumption on how loss is a function of mixture proportions. We find that existing parameterizations of mixing laws can express the true loss-proportion relationship empirically, but the methods themselves often set the mixing law parameters inaccurately, resulting in poor and inconsistent performance. Finally, we leverage the insights from our framework to derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Empirically, Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.28 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.01 test perplexity points.

The rapid acceleration of global urbanization has introduced novel challenges in enhancing urban infrastructure and services. Spatio-temporal data, integrating spatial and temporal dimensions, has emerged as a critical tool for understanding urban phenomena and promoting sustainability. In this context, Federated Learning (FL) has gained prominence as a distributed learning paradigm aligned with the privacy requirements of urban IoT environments. However, integrating traditional and deep learning models into the FL framework poses significant challenges, particularly in capturing complex spatio-temporal dependencies and adapting to diverse urban conditions. To address these challenges, we propose the Federated Local Data-Infused Graph Creation with Node-centric Model Refinement (Fed-LDR) algorithm. Fed-LDR leverages FL and Graph Convolutional Networks (GCN) to enhance spatio-temporal data analysis in urban environments. The algorithm comprises two key modules: (1) the Local Data-Infused Graph Creation (LDIGC) module, which dynamically reconfigures adjacency matrices to reflect evolving spatial relationships within urban environments, and (2) the Node-centric Model Refinement (NoMoR) module, which customizes model parameters for individual urban nodes to accommodate heterogeneity. Evaluations on the PeMSD4 and PeMSD8 datasets demonstrate Fed-LDR's superior performance over six baseline methods. Fed-LDR achieved the lowest Mean Absolute Error (MAE) values of 20.15 and 17.30, and the lowest Root Mean Square Error (RMSE) values of 32.30 and 27.15, respectively, while maintaining a high correlation coefficient of 0.96 across both datasets. Notably, on the PeMSD4 dataset, Fed-LDR reduced MAE and RMSE by up to 81\% and 78\%, respectively, compared to the best-performing baseline FedMedian.

Bilevel Optimization has experienced significant advancements recently with the introduction of new efficient algorithms. Mirroring the success in single-level optimization, stochastic gradient-based algorithms are widely used in bilevel optimization. However, a common limitation in these algorithms is the presumption of independent sampling, which can lead to increased computational costs due to the complicated hyper-gradient formulation of bilevel problems. To address this challenge, we study the example-selection strategy for bilevel optimization in this work. More specifically, we introduce a without-replacement sampling based algorithm which achieves a faster convergence rate compared to its counterparts that rely on independent sampling. Beyond the standard bilevel optimization formulation, we extend our discussion to conditional bilevel optimization and also two special cases: minimax and compositional optimization.

The Gaussian process (GP) is a widely used probabilistic machine learning method for stochastic function approximation, stochastic modeling, and analyzing real-world measurements of nonlinear processes. Unlike many other machine learning methods, GPs include an implicit characterization of uncertainty, making them extremely useful across many areas of science, technology, and engineering. Traditional implementations of GPs involve stationary kernels (also termed covariance functions) that limit their flexibility and exact methods for inference that prevent application to data sets with more than about ten thousand points. Modern approaches to address stationarity assumptions generally fail to accommodate large data sets, while all attempts to address scalability focus on approximating the Gaussian likelihood, which can involve subjectivity and lead to inaccuracies. In this work, we explicitly derive an alternative kernel that can discover and encode both sparsity and nonstationarity. We embed the kernel within a fully Bayesian GP model and leverage high-performance computing resources to enable the analysis of massive data sets. We demonstrate the favorable performance of our novel kernel relative to existing exact and approximate GP methods across a variety of synthetic data examples. Furthermore, we conduct space-time prediction based on more than one million measurements of daily maximum temperature and verify that our results outperform state-of-the-art methods in the Earth sciences. More broadly, having access to exact GPs that use ultra-scalable, sparsity-discovering, nonstationary kernels allows GP methods to truly compete with a wide variety of machine learning methods.

Tabular data plays a pivotal role in various fields, making it a popular format for data manipulation and exchange, particularly on the web. The interpretation, extraction, and processing of tabular information are invaluable for knowledge-intensive applications. Notably, significant efforts have been invested in annotating tabular data with ontologies and entities from background knowledge graphs, a process known as Semantic Table Interpretation (STI). STI automation aids in building knowledge graphs, enriching data, and enhancing web-based question answering. This survey aims to provide a comprehensive overview of the STI landscape. It starts by categorizing approaches using a taxonomy of 31 attributes, allowing for comparisons and evaluations. It also examines available tools, assessing them based on 12 criteria. Furthermore, the survey offers an in-depth analysis of the Gold Standards used for evaluating STI approaches. Finally, it provides practical guidance to help end-users choose the most suitable approach for their specific tasks while also discussing unresolved issues and suggesting potential future research directions.

The joint implementation of Federated learning (FL) and Explainable artificial intelligence (XAI) will allow training models from distributed data and explaining their inner workings while preserving important aspects of privacy. Towards establishing the benefits and tensions associated with their interplay, this scoping review maps those publications that jointly deal with FL and XAI, focusing on publications where an interplay between FL and model interpretability or post-hoc explanations was found. In total, 37 studies met our criteria, with more papers focusing on explanation methods (mainly feature relevance) than on interpretability (mainly algorithmic transparency). Most works used simulated horizontal FL setups involving 10 or fewer data centers. Only one study explicitly and quantitatively analyzed the influence of FL on model explanations, revealing a significant research gap. Aggregation of interpretability metrics across FL nodes created generalized global insights at the expense of node-specific patterns being diluted. 8 papers addressed the benefits of incorporating explanation methods as a component of the FL algorithm. Studies using established FL libraries or following reporting guidelines are a minority. More quantitative research and structured, transparent practices are needed to fully understand their mutual impact and under which conditions it happens.

Human writers plan, then write. For large language models (LLMs) to play a role in longer-form article generation, we must understand the planning steps humans make before writing. We explore one kind of planning, source-selection in news, as a case-study for evaluating plans in long-form generation. We ask: why do specific stories call for specific kinds of sources? We imagine a generative process for story writing where a source-selection schema is first selected by a journalist, and then sources are chosen based on categories in that schema. Learning the article's plan means predicting the schema initially chosen by the journalist. Working with professional journalists, we adapt five existing schemata and introduce three new ones to describe journalistic plans for the inclusion of sources in documents. Then, inspired by Bayesian latent-variable modeling, we develop metrics to select the most likely plan, or schema, underlying a story, which we use to compare schemata. We find that two schemata: stance and social affiliation best explain source plans in most documents. However, other schemata like textual entailment explain source plans in factually rich topics like "Science". Finally, we find we can predict the most suitable schema given just the article's headline with reasonable accuracy. We see this as an important case-study for human planning, and provides a framework and approach for evaluating other kinds of plans. We release a corpora, NewsSources, with annotations for 4M articles.

Despite their power, Transformers face challenges with long sequences due to the quadratic complexity of self-attention. To address this limitation, methods like $k$-Nearest-Neighbor ($k$NN) attention have been introduced [Roy, Saffar, Vaswani, Grangier, 2021] enabling each token to attend to only its $k$ closest tokens. While $k$NN attention has shown empirical success in making Transformers more efficient, its exact approximation guarantees have not been theoretically analyzed. In this work, we establish a theoretical framework for $k$NN attention, reformulating self-attention as expectations over softmax distributions and leveraging lazy Gumbel sampling [Mussmann, Levy, Ermon, 2017] with $k$NN indices for efficient approximation. Building on this framework, we also propose novel sub-quadratic algorithms that approximate self-attention gradients by leveraging efficient sampling techniques, such as Markov Chain-based estimation. Finally, we demonstrate the practical effectiveness of these algorithms through empirical experiments, showcasing their benefits in both training and inference.

Average-link is widely recognized as one of the most popular and effective methods for building hierarchical agglomerative clustering. The available theoretical analyses show that this method has a much better approximation than other popular heuristics, as single-linkage and complete-linkage, regarding variants of Dasgupta's cost function [STOC 2016]. However, these analyses do not separate average-link from a random hierarchy and they are not appealing for metric spaces since every hierarchical clustering has a 1/2 approximation with regard to the variant of Dasgupta's function that is employed for dissimilarity measures [Moseley and Yang 2020]. In this paper, we present a comprehensive study of the performance of average-link in metric spaces, regarding several natural criteria that capture separability and cohesion, and are more interpretable than Dasgupta's cost function and its variants. We also present experimental results with real datasets that, together with our theoretical analyses, suggest that average-link is a better choice than other related methods when both cohesion and separability are important goals.