Language models are increasingly deployed in interactive settings where users reason about facts over time rather than in isolation.In such scenarios, correct behavior requires models to maintain and update implicit temporal assumptions established earlier in a conversation.We study this challenge through the lens of temporal scope stability: the ability to preserve, override, or transfer time-scoped factual context across dialogue turns.We introduce ChronoScope, a large-scale diagnostic benchmark designed to isolate temporal scope behavior in controlled multi-turn interactions, comprising over one million deterministically generated question chains grounded in Wikidata.ChronoScope evaluates whether models can correctly retain inferred temporal scope when follow-up questions omit explicit time references, spanning implicit carryover, explicit scope switching, cross-entity transfer, and longer temporal trajectories.Through extensive evaluation of state-of-the-art language models, we find that temporal scope stability is frequently violated in controlled multiturn settings, with models often drifting toward present-day assumptions despite correct underlying knowledge.These failures intensify with interaction length and persist even under oracle context conditions, revealing a gap between single-turn factual accuracy and coherent temporal reasoning under sequential interaction.We make our dataset and evaluation suite publicly available at https://github. com/yashkumaratri/ChronoScope.

Bryan R Christ, Penelope Molitz, Beau LeBlond, Zachary Gottesman, Jonathan Kropko, Thomas Hartvigsen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Mingtian Tan, Mike A Merrill, Zachary Gottesman, Tim Althoff, David Evans, Thomas Hartvigsen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

In time series editing, we aim to modify some properties of a given time series without altering others. For example, when analyzing a hospital patient's blood pressure, we may add a sudden early drop and observe how it impacts their future, while preserving other conditions. Existing diffusion-based editors rely on rigid, predefined attribute vectors as conditions and produce all-or-nothing edits through sampling. This attribute- and sampling-based approach limits flexibility in condition format and lacks customizable control over editing strength. To overcome these limitations, we introduce Instruction-based Time Series Editing, where users specify intended edits using natural language. This allows users to express a wider range of edits in a more accessible format. We then introduce InstructTime, the first instruction-based time series editor. InstructTime takes in time series and instructions, embeds them into a shared multi-modal representation space, then decodes their embeddings to generate edited time series. By learning a structured multi-modal representation space, we can easily interpolate between embeddings to achieve varying degrees of edit. To handle local and global edits together, we propose multi-resolution encoders. In our experiments, we use synthetic and real datasets and find that InstructTime is a state-of-the-art time series editor: InstructTime achieves high-quality edits with controllable strength, can generalize to unseen instructions, and can be easily adapted to unseen conditions through few-shot learning.

OBJECTIVE: Extracting social determinants of health (SDoHs) from medical notes depends heavily on labor-intensive annotations, which are typically task-specific, hampering reusability and limiting sharing. Here, we introduce SDoH-GPT, a novel framework leveraging few-shot learning large language models (LLMs) to automate the extraction of SDoH from unstructured text, aiming to improve both efficiency and generalizability. MATERIALS AND METHODS: SDoH-GPT is a framework including the few-shot learning LLM methods to extract the SDoH from medical notes and the XGBoost classifiers which continue to classify SDoH using the annotations generated by the few-shot learning LLM methods as training datasets. The unique combination of the few-shot learning LLM methods with XGBoost utilizes the strength of LLMs as great few shot learners and the efficiency of XGBoost when the training dataset is sufficient. Therefore, SDoH-GPT can extract SDoH without relying on extensive medical annotations or costly human intervention. RESULTS: Our approach achieved tenfold and twentyfold reductions in time and cost, respectively, and superior consistency with human annotators measured by Cohen's kappa of up to 0.92. The innovative combination of LLM and XGBoost can ensure high accuracy and computational efficiency while consistently maintaining 0.90+ AUROC scores. DISCUSSION: This study has verified SDoH-GPT on three datasets and highlights the potential of leveraging LLM and XGBoost to revolutionize medical note classification, demonstrating its capability to achieve highly accurate classifications with significantly reduced time and cost. CONCLUSION: The key contribution of this study is the integration of LLM with XGBoost, which enables cost-effective and high quality annotations of SDoH. This research sets the stage for SDoH can be more accessible, scalable, and impactful in driving future healthcare solutions.

Math reasoning is an active area of Large Language Model (LLM) research because it is a hallmark of artificial intelligence and has implications in several domains, including math education. However, few works have explored how math reasoning is encoded within LLM parameters and if it is a skill that can be isolated within models. Doing so could allow targeted intervention to improve math performance without altering non-math behavior and foster understanding of how models encode math reasoning. We introduce Math Neurosurgery (MathNeuro), a computationally efficient method we use to isolate math-specific parameters in LLMs using only forward passes. MathNeuro builds on existing work by using weights and activations to calculate parameter importance, but isolates math-specific parameters by filtering out those important for general language tasks. Through pruning parameters MathNeuro identifies, we delete a LLM’s math reasoning ability without significantly impacting its general language ability. Scaling the identified parameters by a small constant improves a pretrained or instruction-tuned LLM’s performance by 4-17% on GSM8K and 5-35% on MATH while leaving non-math behavior unaltered. MathNeuro is also data efficient: most of its effectiveness holds when identifying math-specific parameters using a single sample. MathNeuro highlights the potential for future work to intervene on math-specific parameters.

Low-bit quantization improves machine learning model efficiency but surprisingly favors undertrained large language models (LLMs).Larger models or those trained on fewer tokens exhibit less quantization-induced degradation (QiD), while smaller, well-trained models face significant performance losses.To gain deeper insights into this trend, we study over 1500+ quantized LLM checkpoints of various sizes and at different training levels (undertrained or fully trained) in a controlled setting, deriving scaling laws for understanding the relationship between QiD and factors: the number of training tokens, model size and bit width.With our derived scaling laws, we propose a novel perspective that we can use QiD to measure an LLM's training levels and determine the number of training tokens required for fully training LLMs of various sizes.Moreover, we use the scaling laws to predict the quantization performance of different-sized LLMs trained with 100 trillion tokens.Our projection shows that the low-bit quantization performance of future models, which are expected to be trained with over 100 trillion tokens, may NOT be desirable.This poses a potential challenge for low-bit quantization in the future and highlights the need for awareness of a model's training level when evaluating lowbit quantization research.To facilitate future research on this problem, we release all the 1500+ quantized checkpoints used in this work at https://huggingface.co/Xu-Ouyang.

Large pre-trained models have achieved notable success across a range of downstream tasks. However, recent research shows that a type of adversarial attack (i.e., backdoor attack) can manipulate the behavior of machine learning models through contaminating their training dataset, posing significant threat in the real-world application of large pre-trained model, especially for those customized models. Therefore, addressing the unique challenges for exploring vulnerability of pre-trained models is of paramount importance. Through empirical studies on the capability for performing backdoor attack in large pre-trained models (e.g., ViT), we find the following unique challenges of attacking large pre-trained models: 1) the inability to manipulate or even access large training datasets, and 2) the substantial computational resources required for training or fine-tuning these models. To address these challenges, we establish new standards for an effective and feasible backdoor attack in the context of large pre-trained models. In line with these standards, we introduce our EDT model, an Efficient, Data-free, Training-free backdoor attack method. Inspired by model editing techniques, EDT injects an editing-based lightweight codebook into the backdoor of large pre-trained models, which replaces the embedding of the poisoned image with the target image without poisoning the training dataset or training the victim model. Our experiments, conducted across various pre-trained models such as ViT, CLIP, BLIP, and stable diffusion, and on downstream tasks including image classification, image captioning, and image generation, demonstrate the effectiveness of our method. Our code is available at https://github.com/donglgcn/Editing/

Sparse Autoencoders (SAEs) provide potentials for uncovering structured, human-interpretable representations in Large Language Models (LLMs), making them a crucial tool for transparent and controllable AI systems.We systematically analyze SAE for interpretable feature extraction from LLMs in safety-critical classification tasks 1 .Our framework evaluates (1) model-layer selection and scaling properties, (2) SAE architectural configurations, including width and pooling strategies, and (3) the effect of binarizing continuous SAE activations.SAE-derived features achieve macro F1 > 0.8, outperforming hidden-state and BoW baselines while demonstrating cross-model transfer from Gemma 2 2B to 9B-IT models.These features generalize in a zero-shot manner to cross-lingual toxicity detection and visual classification tasks.Our analysis highlights the significant impact of pooling strategies and binarization thresholds, showing that binarization offers an efficient alternative to traditional feature selection while maintaining or improving performance.These findings establish new best practices for SAE-based interpretability and enable scalable, transparent deployment of LLMs in real-world applications.

Effective content moderation systems require explicit classification criteria, yet online communities like subreddits often operate with diverse, implicit standards.This work introduces a novel approach to identify and extract these implicit criteria from historical moderation data using an interpretable architecture.We represent moderation criteria as score tables of lexical expressions associated with content removal, enabling systematic comparison across different communities.Our experiments demonstrate that these extracted lexical patterns effectively replicate the performance of neural moderation models while providing transparent insights into decision-making processes.The resulting criteria matrix reveals significant variations in how seemingly shared norms are actually enforced, uncovering previously undocumented moderation patterns including community-specific tolerances for language, features for topical restrictions, and underlying subcategories of the toxic speech classification.latex 1

Ashima Suvarna, Christina A Chance, Karolina Naranjo, Hamid Palangi, Sophie Hao, Thomas Hartvigsen, Saadia Gabriel. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning. Our results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.

Many time series classification tasks, where labels vary over time, are affected by label noise that also varies over time. Such noise can cause label quality to improve, worsen, or periodically change over time. We first propose and formalize temporal label noise, an unstudied problem for sequential classification of time series. In this setting, multiple labels are recorded over time while being corrupted by a time-dependent noise function. We first demonstrate the importance of modeling the temporal nature of the label noise function and how existing methods will consistently underperform. We then propose methods to train noise-tolerant classifiers by estimating the temporal label noise function directly from data. We show that our methods lead to state-of-the-art performance under diverse types of temporal label noise on real-world datasets.

Large multi-modal models inevitably decay over time as facts update and previously learned information becomes outdated. Traditional approaches such as fine-tuning are often impractical for updating these models due to their size and complexity. Instead, direct knowledge editing within the models presents a more viable solution. Current model editing techniques, however, typically overlook the unique influence ranges of different facts, leading to compromised model performance in terms of both generality and locality. To address this issue, we introduce the concept of the generality-locality trade-off in multi-modal model editing. We develop a new model editing dataset named OKEDIT, specifically designed to effectively evaluate this trade-off. Building on this foundation, we propose BalancEdit, a novel method for balanced model editing that dynamically achieves an optimal balance between generality and locality. BalancEdit utilizes a unique mechanism that generates both positive and negative samples for each fact to accurately determine its influence scope and incorporates these insights into the model’s latent space using a discrete, localized codebook of edits, without modifying the underlying model weights. To our knowledge, this is the first approach explicitly addressing the generality-locality trade-off in multi-modal model editing. Our comprehensive results confirm the effectiveness of BalancEdit, demonstrating minimal trade-offs while maintaining robust editing capabilities. Our code and dataset are available at https://github.com/donglgcn/BalancEdit/tree/MMOKVQA.

Keeping large language models factually up-to-date is crucial for deployment, yet costly retraining remains a challenge. Knowledge editing offers a promising alternative, but methods are only tested on small-scale or synthetic edit benchmarks. In this work, we aim to bridge research into lifelong knowledge editing to real-world edits at practically relevant scale. We first introduce WikiBigEdit; a large-scale benchmark of real-world Wikidata edits, built to automatically extend lifelong for future-proof benchmarking. In its first instance, it includes over 500K question-answer pairs for knowledge editing alongside a comprehensive evaluation pipeline. Finally, we use WikiBigEdit to study existing knowledge editing techniques’ ability to incorporate large volumes of real-world facts and contrast their capabilities to generic modification techniques such as retrieval augmentation and continual finetuning to acquire a complete picture of the practical extent of current lifelong knowledge editing.

Discrete diffusion models are a class of generative models that construct sequences by progressively denoising samples from a categorical noise distribution. Beyond their rapidly growing ability to generate coherent natural language, these models present a new and important opportunity to enforce sequence-level constraints, a capability that current autoregressive models cannot natively provide. This paper capitalizes on this opportunity by introducing $\textit{Constrained Discrete Diffusion}$ (CDD), a novel integration of differentiable constraint optimization within the diffusion process to ensure adherence to constraints, logic rules, or safety requirements for generated sequences. Unlike conventional text generators that often rely on post-hoc filtering or model retraining for controllable generation, CDD directly imposes constraints into the discrete diffusion sampling process, resulting in a training-free and effective approach. Experiments in toxicity-controlled text generation, property-constrained molecule design, and instruction-constrained text completion demonstrate that CDD achieves $\textit{zero constraint violations}$ in a diverse array of tasks while preserving fluency, novelty, and coherence, and outperforming autoregressive and existing discrete diffusion approaches.

Offline reinforcement learning in high-dimensional, discrete action spaces is challenging due to the exponential scaling of the joint action space with the number of sub-actions and the complexity of modeling sub-action dependencies. Existing methods either exhaustively evaluate the action space, making them computationally infeasible, or factorize Q-values, failing to represent joint sub-action effects. We propose \textbf{Bra}nch \textbf{V}alue \textbf{E}stimation (BraVE), a value-based method that uses tree-structured action traversal to evaluate a linear number of joint actions while preserving dependency structure. BraVE outperforms prior offline RL methods by up to $20\times$ in environments with over four million actions.

Characterizing a large language model's (LLM's) knowledge of a given question is challenging. As a result, prior work has primarily examined LLM behavior under knowledge conflicts, where the model's internal parametric memory contradicts information in the external context. However, this does not fully reflect how well the model knows the answer to the question. In this paper, we first introduce a taxonomy of five knowledge statuses based on the consistency and correctness of LLM knowledge modes. We then propose KScope, a hierarchical framework of statistical tests that progressively refines hypotheses about knowledge modes and characterizes LLM knowledge into one of these five statuses. We apply KScope to nine LLMs across four datasets and systematically establish: (1) Supporting context narrows knowledge gaps across models. (2) Context features related to difficulty, relevance, and familiarity drive successful knowledge updates. (3) LLMs exhibit similar feature preferences when partially correct or conflicted, but diverge sharply when consistently wrong. (4) Context summarization constrained by our feature analysis, together with enhanced credibility, further improves update effectiveness and generalizes across LLMs.

Vision-language models, like CLIP (Contrastive Language Image Pretraining), are becoming increasingly popular for a wide range of multimodal retrieval tasks. However, prior work has shown that large language and deep vision models can learn historical biases contained in their training sets, leading to perpetuation of stereotypes and potential downstream harm. In this work, we conduct a systematic analysis of the social biases that are present in CLIP, with a focus on the interaction between image and text modalities. We first propose a taxonomy of social biases called So-B-It, which contains 374 words categorized across ten types of bias. Each type can lead to societal harm if associated with a particular demographic group. Using this taxonomy, we examine images retrieved by CLIP from a facial image dataset using each word as part of a prompt. We find that CLIP frequently displays undesirable associations between harmful words and specific demographic groups, such as retrieving mostly pictures of Middle Eastern men when asked to retrieve images of a "terrorist". Finally, we conduct an analysis of the source of such biases, by showing that the same harmful stereotypes are also present in a large image-text dataset used to train CLIP models for examples of biases that we find. Our findings highlight the importance of evaluating and addressing bias in vision-language models, and suggest the need for transparency and fairness-aware curation of large pre-training datasets.

Spike train classification has recently become an important topic in the machine learning community, where each spike train is a binary event sequence with temporal-sparsity of signals of interest and temporal-noise properties. A promising model for it should follow the design principle of performing intensive computation only when signals of interest appear. So such tasks use mainly Spiking Neural Networks (SNNs) due to their consideration of temporal-sparsity of spike trains. However, the basic mechanism of SNNs ignore the temporal-noise issue, which makes them computationally expensive and thus high power consumption for analyzing spike trains on resource-constrained platforms. As an event-driven model, an SNN neuron makes a reaction given any input signals, making it difficult to quickly find signals of interest. In this paper, we introduce an event-attention mechanism that enables SNNs to dynamically highlight useful signals of the original spike trains. To this end, we propose SkipSNN, which extends existing SNN models by learning to mask out noise by skipping membrane potential updates and shortening the effective size of the computational graph. This process is analogous to how people choose to open and close their eyes to filter the information they see. We evaluate SkipSNN on various neuromorphic tasks and demonstrate that it achieves significantly better computational efficiency and classification accuracy than other state-of-the-art SNNs.

Although pre-trained transformers and reprogrammed text-based LLMs have shown strong performance on time series tasks, the best-performing architectures vary widely across tasks, with most models narrowly focused on specific areas, such as time series forecasting. Unifying predictive and generative time series tasks within a single model remains challenging. We introduce UniTS, a unified multi-task time series model that utilizes task tokenization to integrate predictive and generative tasks into a single framework. UniTS employs a modified transformer block to capture universal time series representations, enabling transferability from a heterogeneous, multi-domain pre-training dataset—characterized by diverse dynamic patterns, sampling rates, and temporal scales—to a wide range of downstream datasets with varied task specifications and data domains. Tested on 38 datasets across human activity sensors, healthcare, engineering, and finance, UniTS achieves superior performance compared to 12 forecasting models, 20 classification models, 18 anomaly detection models, and 16 imputation models, including adapted text-based LLMs. UniTS also demonstrates strong few-shot and prompt capabilities when applied to new domains and tasks. In single-task settings, UniTS outperforms competitive task-specialized time series models. Code and datasets are available at https://github.com/mims-harvard/UniTS.

Vision-language (VL) embedding models have been shown to encode biases present in their training data, such as societal biases that prescribe negative characteristics to members of various racial and gender identities. Due to their wide-spread adoption for various tasks ranging from few-shot classification to text-guided image generation, debiasing VL models is crucial. Debiasing approaches that fine-tune the VL model often suffer from catastrophic forgetting. On the other hand, fine-tuning-free methods typically utilize a ``one-size-fits-all" approach that assumes that correlation with the spurious attribute can be explained using a single linear direction across all possible inputs. In this work, we propose a nonlinear, fine-tuning-free approach for VL embedding model debiasing that tailors the debiasing operation to each unique input. This allows for a more flexible debiasing approach. Additionally, we do not require knowledge of the set of inputs a priori to inference time, making our method more appropriate for online tasks such as retrieval and text guided image generation.

Large language models (LLMs) are being applied to time series forecasting. But are language models actually useful for time series? In a series of ablation studies on three recent and popular LLM-based time series forecasting methods, we find that removing the LLM component or replacing it with a basic attention layer does not degrade forecasting performance---in most cases, the results even improve! We also find that despite their significant computational cost, pretrained LLMs do no better than models trained from scratch, do not represent the sequential dependencies in time series, and do not assist in few-shot settings. Additionally, we explore time series encoders and find that patching and attention structures perform similarly to LLM-based forecasters. All resources needed to reproduce our work are available: https://github.com/BennyTMT/LLMsForTimeSeries.

Multi-label classification (MLC), which assigns multiple labels to each instance, is crucial to domains from computer vision to text mining. Conventional methods for MLC require huge amounts of labeled data to capture complex dependencies between labels. However, such labeled datasets are expensive, or even impossible, to acquire. Worse yet, these pre-trained MLC models can only be used for the particular label set covered in the training data. Despite this severe limitation, few methods exist for expanding the set of labels predicted by pre-trained models. Instead, we acquire vast amounts of new labeled data and retrain a new model from scratch. Here, we propose combining the knowledge from multiple pre-trained models (teachers) to train a new student model that covers the union of the labels predicted by this set of teachers. This student supports a broader label set than any one of its teachers without using labeled data. We call this new problem knowledge amalgamation for multi-label classification. Our new method, Adaptive KNowledge Transfer (ANT), trains a student by learning from each teacher’s partial knowledge of label dependencies to infer the global dependencies between all labels across the teachers. We show that ANT succeeds in unifying label dependencies among teachers, outperforming five state-of-the-art methods on eight real-world datasets.

Generative modeling is a powerful technique that involves creating machine learning models capable of creating new data similar to the data it was trained on. Generative Adversarial Networks (GANs) are a leading approach for generative modeling. However, GAN training is known to be a notoriously difficult task. GAN convergence issues are largely caused by the supports of the real and generated distributions being disjoint. To tackle this open problem, we propose a novel GAN pre-training process that effectively aligns the supports of the generated and real data prior to applying traditional adversarial GAN training. The key component of our method, called AlignGAN, is learning a mapping between the input data distribution and a latent representation defined over a hypersphere, regularized by a One Class Classifier. This successfully encourages the generator to produce samples throughout the support of the real data, while not generating samples outside the support. We maintain support alignment through low-bandwidth noise convolutions and additional One Class regularization, leading to continued stable GAN training. We validate our approach against leading stabilization methods on three benchmark datasets, showing AlignGAN routinely produces the best results.

Brain network discovery aims to find nodes and edges from the spatio-temporal signals obtained by neuroimaging data, such as fMRI scans of human brains. Existing methods tend to derive representative or average brain networks, assuming observed signals are generated by only a single brain activity state. However, the human brain usually involves multiple activity states, which jointly determine the brain activities. The brain regions and their connectivity usually exhibit intricate patterns that are difficult to capture with only a single-state network. Recent studies find that brain parcellation and connectivity change according to the brain activity state. We refer to such brain networks as multi-state, and this mixture can help us understand human behavior. Thus, compared to a single-state network, a multi-state network can prevent us from losing crucial information of cognitive brain network. To achieve this, we propose a new model called MNGL (Multi-state Network Graphical Lasso), which successfully models multi-state brain networks by combining CGL (coherent graphical lasso) with GMM (Gaussian Mixture Model). Using both synthetic and real world ADHD-200 fMRI datasets, we demonstrate that MNGL outperforms recent state-of-the-art alternatives by discovering more explanatory and realistic results.

Deployed language models decay over time due to shifting inputs, changing user needs, or emergent world-knowledge gaps. When such problems are identified, we want to make targeted edits while avoiding expensive retraining. However, current model editors, which modify such behaviors of pre-trained models, degrade model performance quickly across multiple, sequential edits. We propose GRACE, a \textit{lifelong} model editing method, which implements spot-fixes on streaming errors of a deployed model, ensuring minimal impact on unrelated inputs. GRACE writes new mappings into a pre-trained model's latent space, creating a discrete, local codebook of edits without altering model weights. This is the first method enabling thousands of sequential edits using only streaming errors. Our experiments on T5, BERT, and GPT models show GRACE's state-of-the-art performance in making and retaining edits, while generalizing to unseen inputs. Our code is available at [github.com/thartvigsen/grace](https://www.github.com/thartvigsen/grace}).

Interpreting time series models is uniquely challenging because it requires identifying both the location of time series signals that drive model predictions and their matching to an interpretable temporal pattern. While explainers from other modalities can be applied to time series, their inductive biases do not transfer well to the inherently challenging interpretation of time series. We present TimeX, a time series consistency model for training explainers. TimeX trains an interpretable surrogate to mimic the behavior of a pretrained time series model. It addresses the issue of model faithfulness by introducing model behavior consistency, a novel formulation that preserves relations in the latent space induced by the pretrained model with relations in the latent space induced by TimeX. TimeX provides discrete attribution maps and, unlike existing interpretability methods, it learns a latent space of explanations that can be used in various ways, such as to provide landmarks to visually aggregate similar explanations and easily recognize temporal patterns. We evaluate TimeX on eight synthetic and real-world datasets and compare its performance against state-of-the-art interpretability methods. We also conduct case studies using physiological time series. Quantitative evaluations demonstrate that TimeX achieves the highest or second-highest performance in every metric compared to baselines across all datasets. Through case studies, we show that the novel components of TimeX show potential for training faithful, interpretable models that capture the behavior of pretrained time series models.

Positive-Unlabeled (PU) learning methods train a classifier to distinguish between the positive and negative classes given only positive and unlabeled data. While traditional PU methods require the labeled positive samples to be an unbiased sample of the positive distribution, in practice the labeled sample is often a biased draw from the true distribution. Prior work shows that if we know the likelihood that each positive instance will be selected for labeling, referred to as the propensity score, then the biased sample can be used for PU learning. Unfortunately, no prior work has been proposed an inference strategy for which the propensity score is identifiable. In this work, we propose two sets of assumptions under which the propensity score can be uniquely determined: one in which no assumption is made on the functional form of the propensity score (requiring assumptions on the data distribution), and the second which loosens the data assumptions while assuming a functional form for the propensity score. We then propose inference strategies for each case. Our empirical study shows that our approach significantly outperforms the state-of-the-art propensity estimation methods on a rich variety of benchmark datasets.

Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, Ece Kamar. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Recurrent Classifier Chains (RCCs) are a leading approach for multi-label classification as they directly model the interdependencies between classes. Unfortunately, existing RCCs assume that every training instance is completely labeled with all its ground truth classes. In practice often only a subset of an instance's labels are annotated, while the annotations for other classes aremissing. RCCs fail in this missing label scenario, predicting many false negatives and potentially missing important classes. In this work, we propose Robust-RCC, the first strategy for tackling this open problem of RCCs failing formulti-label missing-label data. Robust-RCC is a new type of deep recurrent classifier chain empowered to model inter-class relationships essential for predicting thecomplete label set most likely to match the ground truth. The key to Robust-RCC is the design of the Multi Incomplete Label Risk (MILR) function, which we prove to be equal in expectation to the true risk of the ground truth full label set despite being computed from incompletely labeled data. Our experimental study demonstrates that Robust-RCC consistently beats six state-of-of-the-art methods by as much as 30% in predicting the true labels.

Early classification algorithms help users react faster to their machine learning model's predictions. Early warning systems in hospitals, for example, let clinicians improve their patients' outcomes by accurately predicting infections. While early classification systems are advancing rapidly, a major gap remains: existing systems do not consider irregular time series, which have uneven and often-long gaps between their observations. Such series are notoriously pervasive in impactful domains like healthcare. We bridge this gap and study early classification of irregular time series, a new setting for early classifiers that opens doors to more real-world problems. Our solution, Stop&Hop, uses a continuous-time recurrent network to model ongoing irregular time series in real time, while an irregularity-aware halting policy, trained with reinforcement learning, predicts when to stop and classify the streaming series. By taking real-valued step sizes, the halting policy flexibly decides exactly when to stop ongoing series in real time. This way, Stop&Hop seamlessly integrates information contained in the timing of observations, a new and vital source for early classification in this setting, with the time series values to provide early classifications for irregular time series. Using four synthetic and three real-world datasets, we demonstrate that Stop&Hop consistently makes earlier and more-accurate predictions than state-of-the-art alternatives adapted to this new problem. Our code is publicly available at https://github.com/thartvigsen/StopAndHop.

Explainability helps users trust deep learning solutions for time series classification. However, existing explainability methods for multi-class time series classifiers focus on one class at a time, ignoring relationships between the classes. Instead, when a classifier is choosing between many classes, an effective explanation must show what sets the chosen class apart from the rest. We now formalize this notion, studying the open problem of class-specific explainability for deep time series classifiers, a challenging and impactful problem setting. We design a novel explainability method, DEMUX, which learns saliency maps for explaining deep multi-class time series classifiers by adaptively ensuring that its explanation spotlights the regions in an input time series that a model uses specifically to its predicted class. DEMUX adopts a gradient-based approach composed of three interdependent modules that combine to generate consistent, class-specific saliency maps that remain faithful to the classifier’s behavior yet are easily understood by end users. Our experimental study demonstrates that DEMUX outperforms nine state-of-the-art alternatives on five popular datasets when explaining two types of deep time series classifiers. Further, through a case study, we demonstrate that DEMUX’s explanations indeed highlight what separates the predicted class from the others in the eyes of the classifier.

Foodborne illness is a serious but preventable public health problem – with delays in detecting the associated outbreaks resulting in productivity loss, expensive recalls, public safety hazards, and even loss of life. While social media is a promising source for identifying unreported foodborne illnesses, there is a dearth of labeled datasets for developing effective outbreak detection models. To accelerate the development of machine learning-based models for foodborne outbreak detection, we thus present TWEET-FID (TWEET-Foodborne Illness Detection), the first publicly available annotated dataset for multiple foodborne illness incident detection tasks. TWEET-FID collected from Twitter is annotated with three facets: tweet class, entity type, and slot type, with labels produced by experts as well as by crowdsource workers. We introduce several domain tasks leveraging these three facets: text relevance classification (TRC), entity mention detection (EMD), and slot filling (SF). We describe the end-to-end methodology for dataset design, creation, and labeling for supporting model development for these tasks. A comprehensive set of results for these tasks leveraging state-of-the-art single-and multi-task deep learning methods on the TWEET-FID dataset are provided. This dataset opens opportunities for future research in foodborne outbreak detection.

In important domains from video stream analytics to human context recognition, datasets are only partially-labeled. Worse yet, the labels are often applied sequentially, as annotators choose labels frame-by-frame or timestep-by-timestep in sequence. With labels not collected independently, this results in sequential bias in the labeling. Unfortunately, current state-of-the-art methods for partially labeled data are rendered ineffective under sequential bias. In this work, we propose a novel solution to tackling this open sequential bias problem, called DeepSPU. DeepSPU recovers missing labels by constructing a model of the sequentially biased labeling process itself. This labeling model is then learned jointly with the prediction model that infers the missing labels in an iterative training process. Further, we regulate this training using a theoretically-justified cost functions that prevent our model from converging to incorrect but low-cost solution. Our experimental studies demonstrate that DeepSPU consistently outperforms the state-of-the-art methods by over 10% on a rich variety of real-world datasets.

Sequence classification is essential for domains from medical diagnosis to online advertising. In these settings, data are typically proprietary, and annotations are expensive to acquire. Often times, so few annotations are available that training a robust model from scratch is impractical. Recently, knowledge amalgamation (KA) has emerged as a promising strategy for training models without this hard-to-come-by labeled training dataset. To achieve this, KA methods combine the knowledge of multiple pre-trained teacher models (trained on different classification tasks and proprietary datasets) into one student model that becomes an expert on the union of all teachers’ classes. However, we demonstrate that the state-of-the-art solutions fail in the presence of overconfident teachers, which make confident but incorrect predictions for instances from classes upon which they were not trained. Additionally, to-date no work has explored KA for sequence models. Therefore, we propose and then solve the open problem of semi-supervised KA for sequence classification (SKA). Our SKA approach first learns to estimate how trustworthy each teacher is for a given instance, then rescales the predicted probabilities from all teachers to supervise a student model. Our solution overcomes overconfident teachers through careful use of a very small amount of labeled instances. We demonstrate that this approach beats eight state-of-the-art alternatives on four real-world datasets by on average 15% in accuracy with as little as 2% of training data being annotated.

Open set recognition models address the real-world scenario where classes of data unobserved during training are encountered in testing after deployment. Closed set classifiers wrongly attempt to classify instances from an unknown class as belonging to one of the known classes from the training set, which reduces the model’s accuracy. Ideally, these unknown instances should be recognized as such, while known instances should continue to be accurately classified. Unfortunately, state-of-the-art open set methods solve this problem by making restrictive assumptions on the variance and/or boundedness of the distributions of known classes. In this paper, we propose a novel method, Variational Open-Set Recognition (VOSR) that eliminates these assumptions. VOSR incorporates a closed set classifier, an unknown detector, and a novel Structured Gaussian Mixture Variational Autoencoder (SGM-VAE) that guarantees separable class distributions with known variances in its la-tent space. Further, by encouraging a large distance between class-specific distributions, VOSR increases the likelihood that instances from unknown classes lie in low-probability regions and thus are more readily identifiable. In rigorous evaluation, we demonstrate that VOSR outperforms state-of-the-art open set classifiers with up to a 14% F1 score increase in identifying instances from unknown classes in multiple image classification and human activity recognition datasets.

Human-like explanation for text classification is essential for high-impact settings such as healthcare where human rationales are required to support specialists’ decisions. Conventional approaches learn explanations using attention mechanisms to assign heavy weights to words that have a high impact on a model’s prediction. However, such heavily-weighted words often do not reflect human intuition. To advance human rationale, recent studies propose to supervise attention mechanisms assuming access to a huge set of attention labels collected from humans, called human attention maps (HAMs). Unfortunately, acquiring such HAMs for a huge dataset is very tedious, error-prone, and expensive in practice. Thus, we propose the novel problem of text classification with limited human attention supervision. Specifically, we study the learning of human-like attention weights from a dataset in which all documents contain classification labels but only a few documents provide HAMs. To this end, we design a deep learning architecture, HELAS: Human-like Explanation with Limited Attention Supervision to adaptively learn attention weights that focus on words analogous to a human with very limited attention supervision. HELAS effectively unifies joint learning improving both tasks of text classification and humanlike explanation even with only insufficient supervision labels for the latter task. Our experiments show that HELAS generates attention maps similar to real human annotations raising similarity scores up to 22% over state-of-the-art alternatives, even with as little as 2% of the documents having HAMs. It concurrently improves text classification by driving accuracy up to 19% over four state-of-the-art methods.

Explainable classification is essential to high-impact settings where practitioners requireevidence to support their decisions. However, state-of-the-art deep learning models lack transparency in how they make their predictions. One increasingly popular solution is attribution-based explainability, which finds the impact of input features on the model's predictions. While this is popular for computer vision, little has been done to explain deep time series classifiers.In this work, we study this problem and propose PERT, a novel perturbation-based explainability method designed to explain deep classifiers' decisions on time series. PERT extends beyond recent perturbation methods to generate a saliency map that assigns importance values to the timesteps of the instance-of-interest.

Spike train classification is an important problem in many areas such as healthcare and mobile sensing, where each spike train is a high-dimensional time series of binary values. Conventional research on spike train classification mainly focus on developing Spiking Neural Networks (SNNs) under resource-sufficient settings (e.g., on GPU servers). The neurons of the SNNs are usually densely connected in each layer. However, in many real-world applications, we often need to deploy the SNN models on resource-constrained platforms (e.g., mobile devices) to analyze high-dimensional spike train data. The high resource requirement of the densely-connected SNNs can make them hard to deploy on mobile devices. In this paper, we study the problem of energy-efficient SNNs with sparsely-connected neurons. We propose an SNN model with sparse spatio-temporal coding. Our solution is based on the re-parameterization of weights in an SNN and the application of sparsity regularization during optimization. We compare our work with the state-of-the-art SNNs and demonstrate that our sparse SNNs achieve significantly better computational efficiency on both neuromorphic and standard datasets with comparable classification accuracy. Furthermore, compared with densely-connected SNNs, we show that our method has a better capability of generalization on small-size datasets through extensive experiments.

Exact multi-label classification is the task of assigning each datapoint a set of class labels such that the assigned set exactly matches the ground truth. Optimizing for exact multi-label classification is important in domains where missing a single label can be especially costly, such as in object detection for autonomous vehicles or symptom classification for disease diagnosis. Recurrent Classifier Chains (RCCs), a recurrent neural network extension of ensemble-based classifier chains, are the state-of-the-art exact multi-label classification method for maximizing subset accuracy. However, RCCs iteratively predict classes with an unprincipled ordering, and therefore indiscriminately condition class probabilities. These disadvantages make RCCs prone to predicting inaccurate label sets. In this work we propose Recurrent Bayesian Classifier Chains (RBCCs), which learn a Bayesian network of class dependencies and leverage this network in order to condition the prediction of child nodes only on their parents. By conditioning predictions in this way, we perform principled and non-noisy class prediction. We demonstrate the effectiveness of our RBCC method on a variety of real-world multi-label datasets, where we routinely outperform the state of the art methods for exact multi-label classification.

Motivated by human attention, computational attention mechanisms have been designed to help neural networks adjust their focus on specific parts of the input data.While attention mechanisms are claimed to achieve interpretability, little is known about the actual relationships between machine and human attention.In this work, we conduct the first quantitative assessment of human versus computational attention mechanisms for the text classification task.To achieve this, we design and conduct a large-scale crowd-sourcing study to collect human attention maps that encode the parts of a text that humans focus on when conducting text classification.Based on this new resource of human attention dataset for text classification, YELP-HAT, collected on the publicly available YELP dataset, we perform a quantitative comparative analysis of machine attention maps created by deep learning models and human attention maps.Our analysis offers insights into the relationships between human versus machine attention maps along three dimensions: overlap in word selections, distribution over lexical categories, and context-dependency of sentiment polarity.Our findings open promising future research opportunities ranging from supervised attention to the design of human-centric attentionbased explanations.

Publicly available pre-trained word embeddings are rich sources for turning critical high-dimensional representations of huge text data repositories into meaningful compact vectors essential for text mining applications. With many of such pre-trained embedding sources available, each faces limitations in the appropriateness of their language use for the downstream text-mining tasks. Meta-embeddings aim to tackle this ambiguity challenge by fusing multiple embedding sources into one feature space. However, current meta-embedding methods assume vocabularies across sources are similar or even identical; which unfortunately stands in sharp contrast to the fact that many sources barely overlap. Further, these methods encode a meta-embedding for each word by reconstructing its actual embedding values (word-encoder), while valuable information of relationships (distances) among words within each source are not directly considered. In this work, we instead propose a novel relation-encoder learning approach called Similarity-Preserving Meta-Embedding (SimME) that directly integrates word-pair relationships from partially overlapping embedding sources. SimME embeds words such that their similarities are learned from those observed in multiple pre-trained sources. To handle relations between words that are not present in all sources, we introduce maskout, a new loss term, that steers the learning selectively to the sources containing said relations. SimME consistently outperforms state-of-the-art methods by 10% on average and with up to 20% across several core metrics in 4 popular mining tasks on 23 datasets.

Recurrent Neural Networks (RNNs) are the state-of-the-art approach to sequential learning. However, standard RNNs use the same amount of computation to generate their hidden states at each timestep, regardless of the input data. Recent works have begun to tackle this rigid assumption by imposing a priori-determined patterns for updating the states at each step. These approaches could lend insights into the dynamics of RNNs and possibly speed up inference. However, the pre-determined nature of the current update strategies limits their application. To overcome this, we instead design the first fully-learned approach, SA-RNN, that augments any RNN by predicting discrete update patterns at the fine granularity of individual hidden state neurons. This is achieved through the parameterization of a distribution of update-likelihoods driven by the input data. Unlike related methods, our approach imposes no assumptions on the structure of the update patterns. Better yet, our method adapts its update patterns online, allowing different dimensions to be updated conditionally based on the input. To learn which dimensions to update, the model solves a multi-objective optimization problem, maximizing task performance while minimizing the number of updates based on a unified control. Using five publicly-available datasets spanning three sequential learning settings, we demonstrate that our method consistently achieves higher accuracy with fewer updates compared to state-of-the-art alternatives. We also show the benefits of learning to sparsely-update a large hidden state as opposed to densely-update a small hidden state. As an added benefit, our method can be directly applied to a wide variety of models containing RNN architectures.

Early multi-label classification of time series, the assignment of a label set to a time series before the series is entirely observed, is critical for time-sensitive domains such as healthcare. In such cases, waiting too long to classify can render predictions useless, regardless of their accuracy, while predicting prematurely can result in potentially costly erroneous results. When predicting multiple labels (for example, types of infections), dependencies between labels can be learned and leveraged to improve overall accuracy. Together, reliably predicting the correct label set of a time series while observing as few timesteps as possible is challenging because these goals are contradictory in that fewer timesteps often means worse accuracy. To achieve early yet sufficiently accurate predictions, correlations between labels must be accounted for since direct evidence of some labels may only appear late in the series. We design an effective solution to this open problem, the Recurrent Halting Chain (RHC), that for the first time integrates key innovations in both Early and Multi-label Classification into one multi-objective model. RHC uses a recurrent neural network to jointly model raw time series as well as correlations between labels, resulting in a novel order-free classifier chain that tackles this time-sensitive multi-label learning task. Further, RHC employs a reinforcement learning-based halting network to decide at each timestep which, if any, classes should be predicted, learning to build the label set over time. Using two real-world time-sensitive datasets and popular multi-label metrics, we show that RHC outperforms recent alternatives by predicting more-accurate label sets earlier.

In spite of clinical notes in Electronic Health Records (EHR) providing abundant information about patient health, effective modeling of clinical notes remains in its infancy. A patient's clinical notes correspond to a sequence of free-form texts generated by health care professionals over time; with each note in turn containing a sequence of words. Additionally, notes are accompanied by external attributes at multiple layers such as the time at which each note was created (note level) or the demographics of the patient (patient level). Thus, EHR notes correspond to a nested structure of text sequences augmented with external multi-layer attributes. To model this complex problem, we propose an Attributed Hierarchical Attention model, named HAC-RNN, that integrates multiple RNN layers that encode nested sequential notes with contextual and temporal attention layers that are conditioned on the external attributes. While the bottom layer of HAC-RNN is responsible for contextual summarization of the note content, the top layer combs through the entire timeline of notes to focus on those which are most relevant. These attention layers, which are each conditioned on layer-specific hierarchical attributes, allow personalized predictions through inferring patient profiles.We evaluate HAC-RNN using three real-world medical tasks, detecting in-hospital acquired infections and predicting patient mortality using critical care database MIMIC-III. Our results demonstrate that our model significantly outperforms state-of-the-art techniques for all tasks.

Clinical notes correspond to sequences of reports about patients written over time by health-care professionals. With the availability of massive Electronic Health Record (EHR) datasets composed of such clinical notes, machine learning models on these rich text data series are being developed for patient outcome prediction from infection diagnosis to mortality. While current models focus on content in these unstructured clinical notes, we postulate that timing of the medical events that are explained in the notes are equally crucial. We thus propose a novel attention mechanism composed of dual-attention blocks based on a rich diversity of time representations. We then pair this mechanism with an LSTM, resulting in our proposed time-aware recurrent network TEND-LSTM. TEND-LSTM learns an integrated set of attention weights, with the first attention based on the content of the clinical notes and the second based on when the notes were taken. Together, they are combined using a deep-attention network layer. The proposed dual attention mechanism not only learns a function of time incorporating different aspects of the temporal nature of note instances but also automatically finds a balance between how much attention to put on content versus time. We evaluate our model on six medical tasks using data sets from the publicly-available MIMIC III database from the Beth Israel Deaconess Medical Center. Our results demonstrate that TEND-LSTM outperforms state-of-the-art methods by a large margin.

Early classification of time series is the prediction of the class label of a time series before it is observed in its entirety. In time-sensitive domains where information is collected over time it is worth sacrificing some classification accuracy in favor of earlier predictions, ideally early enough for actions to be taken. However, since accuracy and earliness are contradictory objectives, a solution must address this challenge to discover task-dependent trade-offs. We design an early classification model, called EARLIEST, which tackles this multi-objective optimization problem, jointly learning (1) to classify time series and (2) at which timestep to halt and generate this prediction. By learning the objectives together, we achieve a user-controlled balance between these contradictory goals while capturing their natural relationship. Our model consists of the novel pairing of a recurrent discriminator network with a stochastic policy network, with the latter learning a halting-policy as a reinforcement learning task. The learned policy interprets representations generated by the recurrent model and controls its dynamics, sequentially deciding whether or not to request observations from future timesteps. For a rich variety of datasets (four synthetic and three real-world), we demonstrate that EARLIEST consistently out-performs state-of-the-art alternatives in accuracy and earliness while discovering signal locations without supervision.