Time series foundation models (TSFMs) promise to be powerful tools for a wide range of applications. However, their internal representations and learned concepts are still not well understood. In this study, we investigate the structure and redundancy of representations across various TSFMs, examining the self-similarity of model layers within and across different model sizes. This analysis reveals block-like redundancy in the representations, which can be utilized for informed pruning to improve inference speed and efficiency. We also explore the concepts learned by these models, such as periodicity and trends. We demonstrate how conceptual priors can be derived from TSFM representations and leveraged to steer its outputs toward concept-informed predictions. Our work bridges representational analysis from language and vision models to TSFMs, offering new methods for building more computationally efficient and transparent TSFMs.

Time-series and text data are prevalent in healthcare and frequently co-exist, yet they are typically modeled in isolation. Even studies that jointly model time-series and text, do so by converting time-series to images or graphs. We hypothesize that explicitly modeling time-series jointly with text can improve tasks such as summarization and question answering for time-series data, which have received little attention so far. To address this gap, we introduce JoLT to jointly learn desired representations from pre-trained time-series and text models. JoLT utilizes a Querying Transformer (Q-Former) to align the time-series and text representations. Our experiments on a large real-world electrocardiography dataset for medical time-series summarization show that JoLT outperforms state-of-the-art image captioning approaches.

Federated Learning is an effective approach for learning from data distributed across multiple institutions. While most existing studies are aimed at improving predictive accuracy of models, little work has been done to explain knowledge differences between institutions and the benefits of collaboration. Understanding these differences is critical in cross-silo federated learning domains, e.g., in healthcare or banking, where each institution or silo has a different underlying distribution and stakeholders want to understand how their institution compares to their partners. We introduce Prototype-Informed Cross-Silo Router (PICSR) which utilizes a mixture of experts approach to combine local models derived from multiple silos. Furthermore, by computing data similarity to prototypical samples from each silo, we are able to ground the router’s predictions in the underlying dataset distributions. Experiments on a real-world heart disease prediction dataset show that PICSR retains high performance while enabling further explanations on the differences among institutions compared to a single black-box model.

We introduce MOMENT, a family of open-source foundation models for general-purpose time series analysis. Pre-training large models on time series data is challenging due to (1) the absence of a large and cohesive public time series repository, and (2) diverse time series characteristics which make multi-dataset training onerous. Additionally, (3) experimental benchmarks to evaluate these models, especially in scenarios with limited resources, time, and supervision, are still in their nascent stages. To address these challenges, we compile a large and diverse collection of public time series, called the Time series Pile, and systematically tackle time series-specific challenges to unlock large-scale multi-dataset pre-training. Finally, we build on recent work to design a benchmark to evaluate time series foundation models on diverse tasks and datasets in limited supervision settings. Experiments on this benchmark demonstrate the effectiveness of our pre-trained models with minimal data and task-specific fine-tuning. Finally, we present several interesting empirical observations about large pre-trained time series models. Pre-trained models (AutonLab/MOMENT-1-large) and Time Series Pile (AutonLab/Timeseries-PILE) are available on Huggingface.

Machine learning (ML) models are only as good as the data they are trained on. But recent studies have found datasets widely used to train and evaluate ML models, e.g. ImageNet, to have pervasive labeling errors. Erroneous labels on the train set hurt ML models' ability to generalize, and they impact evaluation and model selection using the test set. Consequently, learning in the presence of labeling errors is an active area of research, yet this field lacks a comprehensive benchmark to evaluate these methods. Most of these methods are evaluated on a few computer vision datasets with significant variance in the experimental protocols. With such a large pool of methods and inconsistent evaluation, it is also unclear how ML practitioners can choose the right models to assess label quality in their data. To this end, we propose a benchmarking environment AQuA to rigorously evaluate methods that enable machine learning in the presence of label noise. We also introduce a design space to delineate concrete design choices of label error detection models. We hope that our proposed design space and benchmark enable practitioners to choose the right tools to improve their label quality and that our benchmark enables objective and rigorous evaluation of machine learning tools facing mislabeled data.

Estimation of treatment efficacy of real-world clinical interventions involves working with continuous time-to-event outcomes such as time-to-death, re-hospitalization, or a composite event that may be subject to censoring. Counterfactual reasoning in such scenarios requires decoupling the effects of confounding physiological characteristics that affect baseline survival rates from the effects of the interventions being assessed. In this paper, we present a latent variable approach to model heterogeneous treatment effects by proposing that an individual can belong to one of latent clusters with distinct response characteristics. We show that this latent structure can mediate the base survival rates and help determine the effects of an intervention. We demonstrate the ability of our approach to discover actionable phenotypes of individuals based on their treatment response on multiple large randomized clinical trials originally conducted to assess appropriate treatment strategies to reduce cardiovascular risk.

Problem solving is one of the most important 21st century skills. However, effectively coaching young students in problem solving is challenging because teachers must continuously monitor their cognitive and affective states, and make real-time pedagogical interventions to maximize their learning outcomes. It is an even more challenging task in social environments with limited human coaching resources. To lessen the cognitive load on a teacher and enable affect-sensitive intelligent tutoring, many researchers have investigated automated cognitive and affective detection methods. However, most of the studies use culturally-sensitive indices of affect that are prone to social editing such as facial expressions, and only few studies have explored involuntary dynamic behavioral signals such as gross body movements. In addition, most current methods rely on expensive labelled data from trained annotators for supervised learning. In this paper, we explore a semi-supervised learning framework that can learn low-dimensional representations of involuntary dynamic behavioral signals (mainly gross-body movements) from a modest number of short time series segments. Experiments on a real-world dataset reveal a significant advantage of these representations in discriminating cognitive disequilibrium and flow, as compared to traditional complexity measures from dynamical systems literature, and demonstrate their potential in transferring learned models to previously unseen subjects.

Problem solving is one of the most important 21st century skills. However, effectively coaching young students in problem solving is challenging because teachers must continuously monitor their cognitive and affective states and make real-time pedagogical interventions to maximize students' learning outcomes. It is an even more challenging task in social environments with limited human coaching resources. To lessen the cognitive load on a teacher and enable affect-sensitive intelligent tutoring, many researchers have investigated automated cognitive and affective detection methods. However, most of the studies use culturally-sensitive indices of affect that are prone to social editing such as facial expressions, and only few studies have explored involuntary dynamic behavioral signals such as gross body movements. In addition, most current methods rely on expensive labelled data from trained annotators for supervised learning. In this paper, we explore a semi-supervised learning framework that can learn low-dimensional representations of involuntary dynamic behavioral signals (mainly gross-body movements) from a modest number of short time series segments. Experiments on a real-world dataset reveal a significant utility of these representations in discriminating cognitive disequilibrium and flow and demonstrate their potential in transferring learned models to previously unseen subjects.

Consumer reviews online may contain suggestions useful for improving the target products and services. Mining suggestions is challenging because the field lacks large labelled and balanced datasets. Furthermore, most prior studies have only focused on mining suggestions in a single domain. In this work, we introduce a novel up-sampling technique to address the problem of class imbalance, and propose a multi-task deep learning approach for mining suggestions from multiple domains. Experimental results on a publicly available dataset show that our up-sampling technique coupled with the multi-task framework outperforms state-of-the-art open domain suggestion mining models in terms of the F-1 measure and AUC.

Advances in education technology are enabling tremendous advances in learning at scale. However, they typically assume resources taken for granted in developed countries, including reliable electricity, high-bandwidth Internet access, fast WiFi, powerful computers, sophisticated sensors, and expert technical support to keep it all working. This paper examines these assumptions in the context of a massive test of learning at scale in a developing country. We examine each assumption, how it was broken, and some workarounds used in a 15-month-long independent controlled evaluation of pre- to posttest learning and social-emotional gains by over 2,000 children in 168 villages in Tanzania. We analyze those gains to characterize who gained how much, using test score data, social-emotional measures, and detailed logs from RoboTutor. We quantify the relative impact of pretest scores, literate aspirations, treatment, and usage on learning gains.

Intelligent Tutoring Systems (ITS) have great potential to change the educational landscape by bringing scientifically tested one-to-one tutoring to remote and under-served areas. However, effective ITSs are too complex to perfect. Instead, a practical guiding principle for ITS development and improvement is to fix what’s most broken. In this paper we present SPOT (Statistical Probe of Tutoring): a tool that mines data logged by an Intelligent Tutoring System to identify the ‘hot spots’ most detrimental to its efficiency and effectiveness in terms of its software reliability, usability, task difficulty, student engagement, and other criteria. SPOT uses heuristics and machine learning to discover, characterize, and prioritize such hot spots in order to focus ITS refinement on what matters most. We applied SPOT to data logged by RoboTutor, an ITS that teaches children basic reading, writing and arithmetic.