Haoran Zhang

Yuzhe Yang, Haoran Zhang, Dina Katabi et al.

Machine learning models often perform poorly on subgroups that are underrepresented in the training data. Yet, little is understood on the variation in mechanisms that cause subpopulation shifts, and how algorithms generalize across such diverse shifts at scale. In this work, we provide a fine-grained analysis of subpopulation shift. We first propose a unified framework that dissects and explains common shifts in subgroups. We then establish a comprehensive benchmark of 20 state-of-the-art algorithms evaluated on 12 real-world datasets in vision, language, and healthcare domains. With results obtained from training over 10,000 models, we reveal intriguing observations for future progress in this space. First, existing algorithms only improve subgroup robustness over certain types of shifts but not others. Moreover, while current algorithms rely on group-annotated validation data for model selection, we find that a simple selection criterion based on worst-class accuracy is surprisingly effective even without any group information. Finally, unlike existing works that solely aim to improve worst-group accuracy (WGA), we demonstrate the fundamental tradeoff between WGA and other important metrics, highlighting the need to carefully choose testing metrics. Code and data are available at: https://github.com/YyzHarry/SubpopBench.

Nadia Nahar, Haoran Zhang, Grace Lewis et al. · cmu

Incorporating machine learning (ML) components into software products raises new software-engineering challenges and exacerbates existing challenges. Many researchers have invested significant effort in understanding the challenges of industry practitioners working on building products with ML components, through interviews and surveys with practitioners. With the intention to aggregate and present their collective findings, we conduct a meta-summary study: We collect 50 relevant papers that together interacted with over 4758 practitioners using guidelines for systematic literature reviews. We then collected, grouped, and organized the over 500 mentions of challenges within those papers. We highlight the most commonly reported challenges and hope this meta-summary will be a useful resource for the research community to prioritize research and education in this field.

Haoran Zhang, Natalie Dullerud, Karsten Roth et al.

Deep learning models have reached or surpassed human-level performance in the field of medical imaging, especially in disease diagnosis using chest x-rays. However, prior work has found that such classifiers can exhibit biases in the form of gaps in predictive performance across protected groups. In this paper, we question whether striving to achieve zero disparities in predictive performance (i.e. group fairness) is the appropriate fairness definition in the clinical setting, over minimax fairness, which focuses on maximizing the performance of the worst-case group. We benchmark the performance of nine methods in improving classifier fairness across these two definitions. We find, consistent with prior work on non-clinical data, that methods which strive to achieve better worst-group performance do not outperform simple data balancing. We also find that methods which achieve group fairness do so by worsening performance for all groups. In light of these results, we discuss the utility of fairness definitions in the clinical setting, advocating for an investigation of the bias-inducing mechanisms in the underlying data generating process whenever possible.

Aparna Balagopalan, Haoran Zhang, Kimia Hamidieh et al.

Machine learning models in safety-critical settings like healthcare are often blackboxes: they contain a large number of parameters which are not transparent to users. Post-hoc explainability methods where a simple, human-interpretable model imitates the behavior of these blackbox models are often proposed to help users trust model predictions. In this work, we audit the quality of such explanations for different protected subgroups using real data from four settings in finance, healthcare, college admissions, and the US justice system. Across two different blackbox model architectures and four popular explainability methods, we find that the approximation quality of explanation models, also known as the fidelity, differs significantly between subgroups. We also demonstrate that pairing explainability methods with recent advances in robust machine learning can improve explanation fairness in some settings. However, we highlight the importance of communicating details of non-zero fidelity gaps to users, since a single solution might not exist across all settings. Finally, we discuss the implications of unfair explanation models as a challenging and understudied problem facing the machine learning community.

Haoran Zhang, Harvineet Singh, Marzyeh Ghassemi et al.

Machine learning models frequently experience performance drops under distribution shifts. The underlying cause of such shifts may be multiple simultaneous factors such as changes in data quality, differences in specific covariate distributions, or changes in the relationship between label and features. When a model does fail during deployment, attributing performance change to these factors is critical for the model developer to identify the root cause and take mitigating actions. In this work, we introduce the problem of attributing performance differences between environments to distribution shifts in the underlying data generating mechanisms. We formulate the problem as a cooperative game where the players are distributions. We define the value of a set of distributions to be the change in model performance when only this set of distributions has changed between environments, and derive an importance weighting method for computing the value of an arbitrary set of distributions. The contribution of each distribution to the total performance change is then quantified as its Shapley value. We demonstrate the correctness and utility of our method on synthetic, semi-synthetic, and real-world case studies, showing its effectiveness in attributing performance changes to a wide range of distribution shifts.

Haoran Zhang, Shuanghao Bai, Wanqi Zhou et al.

Source-free domain generalization (SFDG) tackles the challenge of adapting models to unseen target domains without access to source domain data. To deal with this challenging task, recent advances in SFDG have primarily focused on leveraging the text modality of vision-language models such as CLIP. These methods involve developing a transferable linear classifier based on diverse style features extracted from the text and learned prompts or deriving domain-unified text representations from domain banks. However, both style features and domain banks have limitations in capturing comprehensive domain knowledge. In this work, we propose Prompt-Driven Text Adapter (PromptTA) method, which is designed to better capture the distribution of style features and employ resampling to ensure thorough coverage of domain knowledge. To further leverage this rich domain information, we introduce a text adapter that learns from these style features for efficient domain information storage. Extensive experiments conducted on four benchmark datasets demonstrate that PromptTA achieves state-of-the-art performance. The code is available at https://github.com/zhanghr2001/PromptTA.

Haoran Zhang, Junhui Wang

Longitudinal network consists of a sequence of temporal edges among multiple nodes, where the temporal edges are observed in real time. It has become ubiquitous with the rise of online social platform and e-commerce, but largely under-investigated in literature. In this paper, we propose an efficient estimation framework for longitudinal network, leveraging strengths of adaptive network merging, tensor decomposition and point process. It merges neighboring sparse networks so as to enlarge the number of observed edges and reduce estimation variance, whereas the estimation bias introduced by network merging is controlled by exploiting local temporal structures for adaptive network neighborhood. A projected gradient descent algorithm is proposed to facilitate estimation, where the upper bound of the estimation error in each iteration is established. A thorough analysis is conducted to quantify the asymptotic behavior of the proposed method, which shows that it can significantly reduce the estimation error and also provides guideline for network merging under various scenarios. We further demonstrate the advantage of the proposed method through extensive numerical experiments on synthetic datasets and a militarized interstate dispute dataset.

Taylor W. Killian, Haoran Zhang, Thomas Hartvigsen et al. · utoronto

Prevalent in many real-world settings such as healthcare, irregular time series are challenging to formulate predictions from. It is difficult to infer the value of a feature at any given time when observations are sporadic, as it could take on a range of values depending on when it was last observed. To characterize this uncertainty we present EDICT, a strategy that learns an evidential distribution over irregular time series in continuous time. This distribution enables well-calibrated and flexible inference of partially observed features at any time of interest, while expanding uncertainty temporally for sparse, irregular observations. We demonstrate that EDICT attains competitive performance on challenging time series classification tasks and enabling uncertainty-guided inference when encountering noisy data.

Chen Yang, Junzhuo Li, Xinyao Niu et al.

Uncovering early-stage metrics that reflect final model performance is one core principle for large-scale pretraining. The existing scaling law demonstrates the power-law correlation between pretraining loss and training flops, which serves as an important indicator of the current training state for large language models. However, this principle only focuses on the model's compression properties on the training data, resulting in an inconsistency with the ability improvements on the downstream tasks. Some follow-up works attempted to extend the scaling-law to more complex metrics (such as hyperparameters), but still lacked a comprehensive analysis of the dynamic differences among various capabilities during pretraining. To address the aforementioned limitations, this paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. Through this analysis, we confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes, up to 67 billion parameters. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints. This initiative offers valuable resources to the research community and facilitates the verification and exploration of LLM pretraining by open-source researchers. Besides, we provide empirical summaries, including performance comparisons of different models and capabilities, and tuition of key metrics for different training phases. Based on these findings, we provide a more user-friendly strategy for evaluating the optimization state, offering guidance for establishing a stable pretraining process.

Matthew B. A. McDermott, Haoran Zhang, Lasse Hyldig Hansen et al.

In machine learning (ML), a widespread claim is that the area under the precision-recall curve (AUPRC) is a superior metric for model comparison to the area under the receiver operating characteristic (AUROC) for tasks with class imbalance. This paper refutes this notion on two fronts. First, we theoretically characterize the behavior of AUROC and AUPRC in the presence of model mistakes, establishing clearly that AUPRC is not generally superior in cases of class imbalance. We further show that AUPRC can be a harmful metric as it can unduly favor model improvements in subpopulations with more frequent positive labels, heightening algorithmic disparities. Next, we empirically support our theory using experiments on both semi-synthetic and real-world fairness datasets. Prompted by these insights, we conduct a review of over 1.5 million scientific papers to understand the origin of this invalid claim, finding that it is often made without citation, misattributed to papers that do not argue this point, and aggressively over-generalized from source arguments. Our findings represent a dual contribution: a significant technical advancement in understanding the relationship between AUROC and AUPRC and a stark warning about unchecked assumptions in the ML community.

Craig W. Schmidt, Varshini Reddy, Haoran Zhang et al.

Tokenization is a foundational step in natural language processing (NLP) tasks, bridging raw text and language models. Existing tokenization approaches like Byte-Pair Encoding (BPE) originate from the field of data compression, and it has been suggested that the effectiveness of BPE stems from its ability to condense text into a relatively small number of tokens. We test the hypothesis that fewer tokens lead to better downstream performance by introducing PathPiece, a new tokenizer that segments a document's text into the minimum number of tokens for a given vocabulary. Through extensive experimentation we find this hypothesis not to be the case, casting doubt on the understanding of the reasons for effective tokenization. To examine which other factors play a role, we evaluate design decisions across all three phases of tokenization: pre-tokenization, vocabulary construction, and segmentation, offering new insights into the design of effective tokenizers. Specifically, we illustrate the importance of pre-tokenization and the benefits of using BPE to initialize vocabulary construction. We train 64 language models with varying tokenization, ranging in size from 350M to 2.4B parameters, all of which are made publicly available.

Walter Gerych, Haoran Zhang, Kimia Hamidieh et al.

Vision-language model (VLM) embeddings 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. VLMs are being quickly adopted for a variety of tasks ranging from few-shot classification to text-guided image generation, making debiasing VLM embeddings crucial. Debiasing approaches that fine-tune the VLM 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 Bend-VLM, a nonlinear, fine-tuning-free approach for VLM embedding 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, open-set tasks such as retrieval and text guided image generation.

Marie Siew, Haoran Zhang, Jong-Ik Park et al.

Federated learning (FL) enables collaborative learning across multiple clients. In most FL work, all clients train a single learning task. However, the recent proliferation of FL applications may increasingly require multiple FL tasks to be trained simultaneously, sharing clients' computing and communication resources, which we call Multiple-Model Federated Learning (MMFL). Current MMFL algorithms use naive average-based client-task allocation schemes that can lead to unfair performance when FL tasks have heterogeneous difficulty levels, e.g., tasks with larger models may need more rounds and data to train. Just as naively allocating resources to generic computing jobs with heterogeneous resource needs can lead to unfair outcomes, naive allocation of clients to FL tasks can lead to unfairness, with some tasks having excessively long training times, or lower converged accuracies. Furthermore, in the FL setting, since clients are typically not paid for their training effort, we face a further challenge that some clients may not even be willing to train some tasks, e.g., due to high computational costs, which may exacerbate unfairness in training outcomes across tasks. We address both challenges by firstly designing FedFairMMFL, a difficulty-aware algorithm that dynamically allocates clients to tasks in each training round. We provide guarantees on airness and FedFairMMFL's convergence rate. We then propose a novel auction design that incentivizes clients to train multiple tasks, so as to fairly distribute clients' training efforts across the tasks. We show how our fairness-based learning and incentive mechanisms impact training convergence and finally evaluate our algorithm with multiple sets of learning tasks on real world datasets.

Hyewon Jeong, Sarah Jabbour, Yuzhe Yang et al. · uw

The third ML4H symposium was held in person on December 10, 2023, in New Orleans, Louisiana, USA. The symposium included research roundtable sessions to foster discussions between participants and senior researchers on timely and relevant topics for the \ac{ML4H} community. Encouraged by the successful virtual roundtables in the previous year, we organized eleven in-person roundtables and four virtual roundtables at ML4H 2022. The organization of the research roundtables at the conference involved 17 Senior Chairs and 19 Junior Chairs across 11 tables. Each roundtable session included invited senior chairs (with substantial experience in the field), junior chairs (responsible for facilitating the discussion), and attendees from diverse backgrounds with interest in the session's topic. Herein we detail the organization process and compile takeaways from these roundtable discussions, including recent advances, applications, and open challenges for each topic. We conclude with a summary and lessons learned across all roundtables. This document serves as a comprehensive review paper, summarizing the recent advancements in machine learning for healthcare as contributed by foremost researchers in the field.

Haoran Zhang, Quaid Morris, Berk Ustun et al.

Checklists are simple decision aids that are often used to promote safety and reliability in clinical applications. In this paper, we present a method to learn checklists for clinical decision support. We represent predictive checklists as discrete linear classifiers with binary features and unit weights. We then learn globally optimal predictive checklists from data by solving an integer programming problem. Our method allows users to customize checklists to obey complex constraints, including constraints to enforce group fairness and to binarize real-valued features at training time. In addition, it pairs models with an optimality gap that can inform model development and determine the feasibility of learning sufficiently accurate checklists on a given dataset. We pair our method with specialized techniques that speed up its ability to train a predictive checklist that performs well and has a small optimality gap. We benchmark the performance of our method on seven clinical classification problems, and demonstrate its practical benefits by training a short-form checklist for PTSD screening. Our results show that our method can fit simple predictive checklists that perform well and that can easily be customized to obey a rich class of custom constraints.

Sindhu C. M. Gowda, Shalmali Joshi, Haoran Zhang et al.

Machine learning models achieve state-of-the-art performance on many supervised learning tasks. However, prior evidence suggests that these models may learn to rely on shortcut biases or spurious correlations (intuitively, correlations that do not hold in the test as they hold in train) for good predictive performance. Such models cannot be trusted in deployment environments to provide accurate predictions. While viewing the problem from a causal lens is known to be useful, the seamless integration of causation techniques into machine learning pipelines remains cumbersome and expensive. In this work, we study and extend a causal pre-training debiasing technique called causal bootstrapping (CB) under five practical confounded-data generation-acquisition scenarios (with known and unknown confounding). Under these settings, we systematically investigate the effect of confounding bias on deep learning model performance, demonstrating their propensity to rely on shortcut biases when these biases are not properly accounted for. We demonstrate that such a causal pre-training technique can significantly outperform existing base practices to mitigate confounding bias on real-world domain generalization benchmarking tasks. This systematic investigation underlines the importance of accounting for the underlying data-generating mechanisms and fortifying data-preprocessing pipelines with a causal framework to develop methods robust to confounding biases.

Stephen R. Pfohl, Haoran Zhang, Yizhe Xu et al.

Predictive models for clinical outcomes that are accurate on average in a patient population may underperform drastically for some subpopulations, potentially introducing or reinforcing inequities in care access and quality. Model training approaches that aim to maximize worst-case model performance across subpopulations, such as distributionally robust optimization (DRO), attempt to address this problem without introducing additional harms. We conduct a large-scale empirical study of DRO and several variations of standard learning procedures to identify approaches for model development and selection that consistently improve disaggregated and worst-case performance over subpopulations compared to standard approaches for learning predictive models from electronic health records data. In the course of our evaluation, we introduce an extension to DRO approaches that allows for specification of the metric used to assess worst-case performance. We conduct the analysis for models that predict in-hospital mortality, prolonged length of stay, and 30-day readmission for inpatient admissions, and predict in-hospital mortality using intensive care data. We find that, with relatively few exceptions, no approach performs better, for each patient subpopulation examined, than standard learning procedures using the entire training dataset. These results imply that when it is of interest to improve model performance for patient subpopulations beyond what can be achieved with standard practices, it may be necessary to do so via data collection techniques that increase the effective sample size or reduce the level of noise in the prediction problem.

Chandrajit Bajaj, Avik Roy, Haoran Zhang

Variational Autoencoders (VAEs) have been shown to be remarkably effective in recovering model latent spaces for several computer vision tasks. However, currently trained VAEs, for a number of reasons, seem to fall short in learning invariant and equivariant clusters in latent space. Our work focuses on providing solutions to this problem and presents an approach to disentangle equivariance feature maps in a Lie group manifold by enforcing deep, group-invariant learning. Simultaneously implementing a novel separation of semantic and equivariant variables of the latent space representation, we formulate a modified Evidence Lower BOund (ELBO) by using a mixture model pdf like Gaussian mixtures for invariant cluster embeddings that allows superior unsupervised variational clustering. Our experiments show that this model effectively learns to disentangle the invariant and equivariant representations with significant improvements in the learning rate and an observably superior image recognition and canonical state reconstruction compared to the currently best deep learning models.

Imon Banerjee, Ananth Reddy Bhimireddy, John L. Burns et al.

Background: In medical imaging, prior studies have demonstrated disparate AI performance by race, yet there is no known correlation for race on medical imaging that would be obvious to the human expert interpreting the images. Methods: Using private and public datasets we evaluate: A) performance quantification of deep learning models to detect race from medical images, including the ability of these models to generalize to external environments and across multiple imaging modalities, B) assessment of possible confounding anatomic and phenotype population features, such as disease distribution and body habitus as predictors of race, and C) investigation into the underlying mechanism by which AI models can recognize race. Findings: Standard deep learning models can be trained to predict race from medical images with high performance across multiple imaging modalities. Our findings hold under external validation conditions, as well as when models are optimized to perform clinically motivated tasks. We demonstrate this detection is not due to trivial proxies or imaging-related surrogate covariates for race, such as underlying disease distribution. Finally, we show that performance persists over all anatomical regions and frequency spectrum of the images suggesting that mitigation efforts will be challenging and demand further study. Interpretation: We emphasize that model ability to predict self-reported race is itself not the issue of importance. However, our findings that AI can trivially predict self-reported race -- even from corrupted, cropped, and noised medical images -- in a setting where clinical experts cannot, creates an enormous risk for all model deployments in medical imaging: if an AI model secretly used its knowledge of self-reported race to misclassify all Black patients, radiologists would not be able to tell using the same data the model has access to.

Haoran Zhang, Natalie Dullerud, Laleh Seyyed-Kalantari et al.

Clinical machine learning models experience significantly degraded performance in datasets not seen during training, e.g., new hospitals or populations. Recent developments in domain generalization offer a promising solution to this problem by creating models that learn invariances across environments. In this work, we benchmark the performance of eight domain generalization methods on multi-site clinical time series and medical imaging data. We introduce a framework to induce synthetic but realistic domain shifts and sampling bias to stress-test these methods over existing non-healthcare benchmarks. We find that current domain generalization methods do not consistently achieve significant gains in out-of-distribution performance over empirical risk minimization on real-world medical imaging data, in line with prior work on general imaging datasets. However, a subset of realistic induced-shift scenarios in clinical time series data do exhibit limited performance gains. We characterize these scenarios in detail, and recommend best practices for domain generalization in the clinical setting.

Taylor W. Killian, Haoran Zhang, Jayakumar Subramanian et al.

Reinforcement Learning (RL) has recently been applied to sequential estimation and prediction problems identifying and developing hypothetical treatment strategies for septic patients, with a particular focus on offline learning with observational data. In practice, successful RL relies on informative latent states derived from sequential observations to develop optimal treatment strategies. To date, how best to construct such states in a healthcare setting is an open question. In this paper, we perform an empirical study of several information encoding architectures using data from septic patients in the MIMIC-III dataset to form representations of a patient state. We evaluate the impact of representation dimension, correlations with established acuity scores, and the treatment policies derived from them. We find that sequentially formed state representations facilitate effective policy learning in batch settings, validating a more thoughtful approach to representation learning that remains faithful to the sequential and partial nature of healthcare data.

Haoran Zhang, Diane Litman

While automated essay scoring (AES) can reliably grade essays at scale, automated writing evaluation (AWE) additionally provides formative feedback to guide essay revision. However, a neural AES typically does not provide useful feature representations for supporting AWE. This paper presents a method for linking AWE and neural AES, by extracting Topical Components (TCs) representing evidence from a source text using the intermediate output of attention layers. We evaluate performance using a feature-based AES requiring TCs. Results show that performance is comparable whether using automatically or manually constructed TCs for 1) representing essays as rubric-based features, 2) grading essays.

Qingyun Wang, Manling Li, Xuan Wang et al.

To combat COVID-19, both clinicians and scientists need to digest vast amounts of relevant biomedical knowledge in scientific literature to understand the disease mechanism and related biological functions. We have developed a novel and comprehensive knowledge discovery framework, COVID-KG to extract fine-grained multimedia knowledge elements (entities and their visual chemical structures, relations, and events) from scientific literature. We then exploit the constructed multimedia knowledge graphs (KGs) for question answering and report generation, using drug repurposing as a case study. Our framework also provides detailed contextual sentences, subfigures, and knowledge subgraphs as evidence.

Haoran Zhang, Amy X. Lu, Mohamed Abdalla et al.

In this work, we examine the extent to which embeddings may encode marginalized populations differently, and how this may lead to a perpetuation of biases and worsened performance on clinical tasks. We pretrain deep embedding models (BERT) on medical notes from the MIMIC-III hospital dataset, and quantify potential disparities using two approaches. First, we identify dangerous latent relationships that are captured by the contextual word embeddings using a fill-in-the-blank method with text from real clinical notes and a log probability bias score quantification. Second, we evaluate performance gaps across different definitions of fairness on over 50 downstream clinical prediction tasks that include detection of acute and chronic conditions. We find that classifiers trained from BERT representations exhibit statistically significant differences in performance, often favoring the majority group with regards to gender, language, ethnicity, and insurance status. Finally, we explore shortcomings of using adversarial debiasing to obfuscate subgroup information in contextual word embeddings, and recommend best practices for such deep embedding models in clinical settings.

Haoran Zhang, Diane Litman

This paper presents an investigation of using a co-attention based neural network for source-dependent essay scoring. We use a co-attention mechanism to help the model learn the importance of each part of the essay more accurately. Also, this paper shows that the co-attention based neural network model provides reliable score prediction of source-dependent responses. We evaluate our model on two source-dependent response corpora. Results show that our model outperforms the baseline on both corpora. We also show that the attention of the model is similar to the expert opinions with examples.

Haoran Zhang, Diane Litman

Manually grading the Response to Text Assessment (RTA) is labor intensive. Therefore, an automatic method is being developed for scoring analytical writing when the RTA is administered in large numbers of classrooms. Our long-term goal is to also use this scoring method to provide formative feedback to students and teachers about students' writing quality. As a first step towards this goal, interpretable features for automatically scoring the evidence rubric of the RTA have been developed. In this paper, we present a simple but promising method for improving evidence scoring by employing the word embedding model. We evaluate our method on corpora of responses written by upper elementary students.