Graph neural networks (GNNs) are the dominant approach to solving machine learning problems defined over graphs. Despite much theoretical and empirical work in recent years, our understanding of finer-grained aspects of architectural design for GNNs remains impoverished. In this paper, we consider the benefits of architectures that maintain and update edge embeddings. On the theoretical front, under a suitable computational abstraction for a layer in the model, as well as memory constraints on the embeddings, we show that there are natural tasks on graphical models for which architectures leveraging edge embeddings can be much shallower. Our techniques are inspired by results on time-space tradeoffs in theoretical computer science. Empirically, we show architectures that maintain edge embeddings almost always improve on their node-based counterparts---frequently significantly so in topologies that have "hub" nodes.

Predictions and generations from large language models are increasingly being explored as an aid in limited data regimes, such as in computational social science and human subjects research. While prior technical work has mainly explored the potential to use model-predicted labels for unlabeled data in a principled manner, there is increasing interest in using large language models to generate entirely new synthetic samples (e.g., synthetic simulations), such as in responses to surveys. However, it remains unclear by what means practitioners can combine such data with real data and yet produce statistically valid conclusions upon them. In this paper, we introduce a new estimator based on generalized method of moments, providing a hyperparameter-free solution with strong theoretical guarantees to address this challenge. Intriguingly, we find that interactions between the moment residuals of synthetic data and those of real data (i.e., when they are predictive of each other) can greatly improve estimates of the target parameter. We validate the finite-sample performance of our estimator across different tasks in computational social science applications, demonstrating large empirical gains.

Robust unlearning is crucial for safely deploying large language models (LLMs) in environments where data privacy, model safety, and regulatory compliance must be ensured. Yet the task is inherently challenging, partly due to difficulties in reliably measuring whether unlearning has truly occurred. Moreover, fragmentation in current methodologies and inconsistent evaluation metrics hinder comparative analysis and reproducibility. To unify and accelerate research efforts, we introduce OpenUnlearning, a standardized and extensible framework designed explicitly for benchmarking both LLM unlearning methods and metrics. OpenUnlearning integrates 13 state-of-the-art unlearning algorithms and 16 diverse evaluations across 3 leading benchmarks (TOFU, MUSE, and WMDP) and also enables analyses of forgetting behaviors across 450+ publicly released checkpoints. Leveraging OpenUnlearning, we propose a novel meta-evaluation benchmark focused specifically on assessing the faithfulness and robustness of evaluation metrics themselves. We also benchmark diverse unlearning methods and provide a comparative analysis against an extensive evaluation suite. Overall, we establish a clear, community-driven pathway toward rigorous development in LLM unlearning research.

As large language models (LLMs) are increasingly deployed in high-stakes settings, the risk of generating harmful or toxic content remains a central challenge. Post-hoc alignment methods are brittle: once unsafe patterns are learned during pretraining, they are hard to remove. In this work, we present a data-centric pretraining framework that builds safety into the model from the start. Our framework consists of four key steps: (i) Safety Filtering: building a safety classifier to classify webdata into safe and unsafe categories; (ii) Safety Rephrasing: we recontextualize unsafe webdata into safer narratives; (iii) Native Refusal: we synthetically generate pretraining datasets that actively teach models to refuse on unsafe content and the moral reasoning behind it, and (iv) Harmfulness-Tag annotated pretraining: we flag unsafe content during pretraining using a special token, and use it to steer models away from unsafe generations at inference-time. Our safety-pretrained models reduce attack success rates from 38.8% to 8.4% on standard LLM safety benchmarks with no performance degradation on general tasks.

Recent advances in general-purpose AI underscore the urgent need to align AI systems with human goals and values. Yet, the lack of a clear, shared understanding of what constitutes "alignment" limits meaningful progress and cross-disciplinary collaboration. In this position paper, we argue that the research community should explicitly define and critically reflect on "alignment" to account for the bidirectional and dynamic relationship between humans and AI. Through a systematic review of over 400 papers spanning HCI, NLP, ML, and more, we examine how alignment is currently defined and operationalized. Building on this analysis, we introduce the Bidirectional Human-AI Alignment framework, which not only incorporates traditional efforts to align AI with human values but also introduces the critical, underexplored dimension of aligning humans with AI – supporting cognitive, behavioral, and societal adaptation to rapidly advancing AI technologies. Our findings reveal significant gaps in current literature, especially in long-term interaction design, human value modeling, and mutual understanding. We conclude with three central challenges and actionable recommendations to guide future research toward more nuanced, reciprocal, and human-AI alignment approaches.

The primary practice of healthcare artificial intelligence (AI) starts with model development, often using state-of-the-art AI, retrospectively evaluated using metrics lifted from the AI literature like AUROC and DICE score. However, good performance on these metrics may not translate to improved clinical outcomes. Instead, we argue for a better development pipeline constructed by working backward from the end goal of positively impacting clinically relevant outcomes using AI, leading to considerations of causality in model development and validation, and subsequently a better development pipeline. Healthcare AI should be "actionable," and the change in actions induced by AI should improve outcomes. Quantifying the effect of changes in actions on outcomes is causal inference. The development, evaluation, and validation of healthcare AI should therefore account for the causal effect of intervening with the AI on clinically relevant outcomes. Using a causal lens, we make recommendations for key stakeholders at various stages of the healthcare AI pipeline. Our recommendations aim to increase the positive impact of AI on clinical outcomes.

Recent advancements in large language models (LLMs) have considerably advanced the capabilities of summarization systems.However, they continue to face concerns about hallucination.While prior work has evaluated LLMs extensively in news domains, most evaluation of dialogue summarization has focused on BART-based models, leaving a gap in our understanding of their faithfulness.Our work benchmarks the faithfulness of LLMs for dialogue summarization, using human annotations and focusing on identifying and categorizing span-level inconsistencies.Specifically, we focus on two prominent LLMs: GPT-4 and Alpaca-13B.Our evaluation reveals subtleties as to what constitutes a hallucination: LLMs often generate plausible inferences, supported by circumstantial evidence in the conversation, that lack direct evidence, a pattern that is less prevalent in older models.We propose a refined taxonomy of errors, coining the category of "Circumstantial Inference" to bucket these LLM behaviors.Using our taxonomy, we compare the behavioral differences between LLMs and older fine-tuned models.Additionally, we systematically assess the efficacy of automatic error detection methods on LLM summaries and find that they struggle to detect these nuanced errors.To address this, we introduce two prompt-based approaches for fine-grained error detection that outperform existing metrics, particularly for identifying "Circumstantial Inference." 1

In response to rising concerns surrounding the safety, security, and trustworthiness of Generative AI (GenAI) models, practitioners and regulators alike have pointed to AI red-teaming as a key component of their strategies for identifying and mitigating these risks. However, despite AI red-teaming’s central role in policy discussions and corporate messaging, significant questions remain about what precisely it means, what role it can play in regulation, and how it relates to conventional red-teaming practices as originally conceived in the field of cybersecurity. In this work, we identify recent cases of red-teaming activities in the AI industry and conduct an extensive survey of relevant research literature to characterize the scope, structure, and criteria for AI red-teaming practices. Our analysis reveals that prior methods and practices of AI red-teaming diverge along several axes, including the purpose of the activity (which is often vague), the artifact under evaluation, the setting in which the activity is conducted (e.g., actors, resources, and methods), and the resulting decisions it informs (e.g., reporting, disclosure, and mitigation). In light of our findings, we argue that while red-teaming may be a valuable big-tent idea for characterizing GenAI harm mitigations, and that industry may effectively apply red-teaming and other strategies behind closed doors to safeguard AI, gestures towards red-teaming (based on public definitions) as a panacea for every possible risk verge on security theater. To move toward a more robust toolbox of evaluations for generative AI, we synthesize our recommendations into a question bank meant to guide and scaffold future AI red-teaming practices.

A fundamental problem in decision-making systems is the presence of inequity along demographic lines. However, inequity can be difficult to quantify, particularly if our notion of equity relies on hard-to-measure notions like risk (e.g., equal access to treatment for those who would die without it). Auditing such inequity requires accurate measurements of individual risk, which is difficult to estimate in the realistic setting of unobserved confounding. In the case that these unobservables “explain” an apparent disparity, we may understate or overstate inequity. In this paper, we show that one can still give informative bounds on allocation rates among high-risk individuals, even while relaxing or (surprisingly) even when eliminating the assumption that all relevant risk factors are observed. We utilize the fact that in many real-world settings (e.g., the introduction of a novel treatment) we have data from a period prior to any allocation, to derive unbiased estimates of risk. We apply our framework to a real-world setting of Paxlovid allocation to COVID-19 patients, finding that observed racial inequity cannot be explained by unobserved confounders of the same strength as important observed covariates.

In standard reinforcement learning setups, the agent receives observations and performs actions at evenly spaced intervals. However, in many real-world settings, observations are expensive, forcing agents to commit to courses of action for designated periods of time. Consider that doctors, after each visit, typically set not only a treatment plan but also a follow-up date at which that plan might be revised. In this work, we formalize the setup of timing-as-an-action. Through theoretical analysis in the tabular setting, we show that while the choice of delay intervals could be naively folded in as part of a composite action, these actions have a special structure and handling them intelligently yields statistical advantages. Taking a model-based perspective, these gains owe to the fact that delay actions do not add any parameters to the underlying model. For model estimation, we provide provable sample-efficiency improvements, and our experiments demonstrate empirical improvements in both healthcare simulators and classical reinforcement learning environments.

Simple, sufficient explanations furnished by short decision lists can be useful for guiding stakeholder actions. Unfortunately, this transparency can come at the expense of the higher accuracy enjoyed by black box methods, like deep nets. To date, practitioners typically either (i) insist on the simpler model, forsaking accuracy; or (ii) insist on maximizing accuracy, settling for post-hoc explanations of dubious faithfulness. In this paper, we propose a hybrid partially interpretable model that represents a compromise between the two extremes. In our setup, each input is first processed by a decision list that can either execute a decision or abstain, handing off authority to the opaque model. The key to optimizing the decision list is to optimally trade off the accuracy of the composite system against coverage (the fraction of the population that receives explanations). We contribute a new principled algorithm for constructing partially interpretable decision lists, providing theoretical guarantees addressing both interpretability and accuracy. As an instance of our result, we prove that when the optimal decision list has length $k$, coverage $c$, and $b$ mistakes, our algorithm will generate a decision list that has length no greater than $4k$, coverage at least $c/2$, and makes at most $4b$ mistakes. Finally, we empirically validate the effectiveness of the new model.

Vision-language models (VLMs) are trained for thousands of GPU hours on carefully selected subsets of massive web scrapes. For instance, the LAION public dataset retained only about 10% of the total crawled data. In recent times, data curation has gained prominence with several works developing strategies to retain ‘high-quality’ subsets of ‘raw’ scraped data. However, these strategies are typically developed agnostic to the available compute for training. In this paper, we demonstrate that making filtering decisions independent of training compute is often suboptimal—well-curated data rapidly loses its utility when repeated, eventually decreasing below the utility of ‘unseen’ but ‘lower-quality’ data. While past research in neural scaling laws has considered web data to be homogenous, real data is not. Our work bridges this important gap in the literature by developing scaling laws that characterize the differing ‘utility’ of various data subsets, and accounting for how this diminishes for a data point at its ‘nth’ repetition. Our key message is that data curation can not be agnostic of the total compute a model will be trained for. Even without ever jointly training on multiple data buckets, our scaling laws enable us to estimate model performance under this dynamic trade-off between quality and repetition. This allows us to curate the best possible pool for achieving top performance on Datacomp at various compute budgets, carving out a pareto-frontier for data curation.

Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pretraining on publicly available biomedical corpora.These works typically claim that such domain-adaptive pretraining (DAPT) improves performance on downstream medical tasks, such as answering medical licensing exam questions.In this paper, we compare seven public "medical" LLMs and two VLMs against their corresponding base models, arriving at a different conclusion: all medical VLMs and nearly all medical LLMs fail to consistently improve over their base models in the zero-/few-shot prompting regime for medical question-answering (QA) tasks.For instance, across the tasks and model pairs we consider in the 3-shot setting, medical LLMs only outperform their base models in 12.1% of cases, reach a (statistical) tie in 49.8% of cases, and are significantly worse than their base models in the remaining 38.2% of cases.Our conclusions are based on (i) comparing each medical model head-to-head, directly against the corresponding base model; (ii) optimizing the prompts for each model separately; and (iii) accounting for statistical uncertainty in comparisons.While these basic practices are not consistently adopted in the literature, our ablations show that they substantially impact conclusions.Our findings suggest that state-of-the-art generaldomain models may already exhibit strong medical knowledge and reasoning capabilities, and offer recommendations to strengthen the conclusions of future studies.Table 1: Summary of open-source autoregressive VLM and LLM pairs used for evaluation.

Real-world deployment of machine learning models is challenging because data evolves over time. While no model can work when data evolves in an arbitrary fashion, if there is some pattern to these changes, we might be able to design methods to address it. This paper addresses situations when data evolves gradually. We introduce a time-varying propensity score that can detect gradual shifts in the distribution of data which allows us to selectively sample past data to update the model---not just similar data from the past like that of a standard propensity score but also data that evolved in a similar fashion in the past. The time-varying propensity score is quite general: we demonstrate different ways of implementing it and evaluate it on a variety of problems ranging from supervised learning (e.g., image classification problems) where data undergoes a sequence of gradual shifts, to reinforcement learning tasks (e.g., robotic manipulation and continuous control) where data shifts as the policy or the task changes.

Large web-crawled multimodal datasets have powered a slew of new methods for learning general-purpose visual representations, advancing the state of the art in computer vision and revolutionizing zero- and few-shot recognition. One crucial decision facing practitioners is how, if at all, to curate these ever-larger datasets. For example, the creators of the LAION-5B dataset chose to retain only image-caption pairs whose CLIP similarity score exceeded a designated threshold. In this paper, we propose a new state-of-the-art data filtering approach motivated by our observation that nearly $40\%$ of LAION's images contain text that overlaps significantly with the caption. Intuitively, such data could be wasteful as it incentivizes models to perform optical character recognition rather than learning visual features. However, naively removing all such data could also be wasteful, as it throws away images that contain visual features (in addition to overlapping text). Our simple and scalable approach, T-MARS (Text Masking and Re-Scoring), filters out only those pairs where the text dominates the remaining visual features---by first masking out the text and then filtering out those with a low CLIP similarity score of the masked image with original captions. Experimentally, T-MARS is the top ranked approach on Imagenet at ``medium scale'' of DataComp (a data filtering benchmark), and outperforms CLIP filtering by a margin of $6.5\%$ on ImageNet and $4.7\%$ on VTAB. Additionally, we show that the accuracy gains enjoyed by T-MARS linearly increase as data and compute are scaled exponentially.

Discovery of causal relationships is a fundamental goal of science and vital for sound decision making. As such, there has been considerable interest in causal discovery methods with provable guarantees. Existing works have thus far largely focused on discovery under hard intervention and infinite-samples, in which intervening on a node readily reveals the orientation of every edge incident to the node. This setup however overlooks the stochasticity inherent in real-world, finite-sample settings. Our work takes a step towards studying finite-sample causal discovery, wherein multiple interventions on a node are now needed for edge orientation. In this work, we study the canonical setup in theoretical causal discovery literature, where one assumes causal sufficiency and access to the graph skeleton. Our key observation is that discovery may be viewed as structured, multiple testing, and we develop a novel testing framework to this end. Crucially, our framework allows for anytime valid testing as multiple tests are needed to conclude an edge orientation. It also allows for flexible combination of structured test-statistics (enabling one to use Meek rules to propagate edge orientation) as well as robust testing. Through empirical simulations, we confirm the usefulness of our framework. In closing, using this testing framework, we show how one may efficiently verify graph structure by drawing a connection to multi-constraint bandits and designing a novel algorithm to this end.

Colloquially speaking, image generation models based upon diffusion processes are frequently said to exhibit ''hallucinations'' samples that could never occur in the training data. But where do such hallucinations come from? In this paper, we study a particular failure mode in diffusion models, which we term ***mode interpolation***. Specifically, we find that diffusion models smoothly ``interpolate'' between nearby data modes in the training set, to generate samples that are completely outside the support of the original training distribution; this phenomenon leads diffusion models to generate artifacts that never existed in real data (i.e., hallucinations). We systematically study the reasons for, and the manifestation of this phenomenon. Through experiments on 1D and 2D Gaussians, we show how a discontinuous loss landscape in the diffusion model's decoder leads to a region where any smooth approximation will cause such hallucinations. Through experiments on artificial datasets with various shapes, we show how hallucination leads to the generation of combinations of shapes that never existed. We extend the validity of mode interpolation in real-world datasets by explaining the unexpected generation of images with additional or missing fingers similar to those produced by popular text-to-image generative models. Finally, we show that diffusion models in fact ***know*** when they go out of support and hallucinate. This is captured by the high variance in the trajectory of the generated sample towards the final few backward sampling process. Using a simple metric to capture this variance, we can remove over 95\% of hallucinations at generation time. We conclude our exploration by showing the implications of such hallucination (and its removal) on the collapse (and stabilization) of recursive training on synthetic data with experiments on datasets like MNIST .

Fairness metrics are a core tool in the fair machine learning literature (FairML),used to determine that ML models are, in some sense, “fair.” Real-world data,however, are typically plagued by various measurement biases and other violatedassumptions, which can render fairness assessments meaningless. We adapt toolsfrom causal sensitivity analysis to the FairML context, providing a general frame-work which (1) accommodates effectively any combination of fairness metric andbias that can be posed in the “oblivious setting”; (2) allows researchers to inves-tigate combinations of biases, resulting in non-linear sensitivity; and (3) enablesflexible encoding of domain-specific constraints and assumptions. Employing thisframework, we analyze the sensitivity of the most common parity metrics under 3varieties of classifier across 14 canonical fairness datasets. Our analysis reveals thestriking fragility of fairness assessments to even minor dataset biases. We show thatcausal sensitivity analysis provides a powerful and necessary toolkit for gaugingthe informativeness of parity metric evaluations. Our repository is \href{https://github.com/Jakefawkes/fragile_fair}{available here}.

Trained models are often composed with post-hoc transforms such as temperature scaling (TS), ensembling and stochastic weight averaging (SWA) to improve performance, robustness, uncertainty estimation, etc. However, such transforms are typically applied only after the base models have already been finalized by standard means. In this paper, we challenge this practice with an extensive empirical study. In particular, we demonstrate a phenomenon that we call post-hoc reversal, where performance trends are reversed after applying post-hoc transforms. This phenomenon is especially prominent in high-noise settings. For example, while base models overfit badly early in training, both ensembling and SWA favor base models trained for more epochs. Post-hoc reversal can also prevent the appearance of double descent and mitigate mismatches between test loss and test error seen in base models. Preliminary analyses suggest that these transforms induce reversal by suppressing the influence of mislabeled examples, exploiting differences in their learning dynamics from those of clean examples. Based on our findings, we propose post-hoc selection, a simple technique whereby post-hoc metrics inform model development decisions such as early stopping, checkpointing, and broader hyperparameter choices. Our experiments span real-world vision, language, tabular and graph datasets. On an LLM instruction tuning dataset, post-hoc selection results in >1.5x MMLU improvement compared to naive selection.

Large language models (LLMs) trained on web-scale datasets raise substantial concerns regarding permissible data usage. One major question is whether these models "memorize" all their training data or they integrate many data sources in some way more akin to how a human would learn and synthesize information. The answer hinges, to a large degree, on \emph{how we define memorization.} In this work, we propose the Adversarial Compression Ratio (ACR) as a metric for assessing memorization in LLMs. A given string from the training data is considered memorized if it can be elicited by a prompt (much) shorter than the string itself---in other words, if these strings can be ``compressed'' with the model by computing adversarial prompts of fewer tokens. The ACR overcomes the limitations of existing notions of memorization by (i) offering an adversarial view of measuring memorization, especially for monitoring unlearning and compliance; and (ii) allowing for the flexibility to measure memorization for arbitrary strings at a reasonably low compute. Our definition serves as a practical tool for determining when model owners may be violating terms around data usage, providing a potential legal tool and a critical lens through which to address such scenarios.

A key source of complexity in next-generation AI models is the size of model outputs, making it time-consuming to parse and provide reliable feedback on. To ensure such models are aligned, we will need to bolster our understanding of scalable oversight and how to scale up human feedback. To this end, we study the challenges of scalable oversight in the context of goal-conditioned hierarchical reinforcement learning. Hierarchical structure is a promising entrypoint into studying how to scale up human feedback, which in this work we assume can only be provided for model outputs below a threshold size. In the cardinal feedback setting, we develop an apt sub-MDP reward and algorithm that allows us to acquire and scale up low-level feedback for learning with sublinear regret. In the ordinal feedback setting, we show the necessity of both high- and low-level feedback, and develop a hierarchical experimental design algorithm that efficiently acquires both types of feedback for learning. Altogether, our work aims to consolidate the foundations of scalable oversight, formalizing and studying the various challenges thereof.

With artificial intelligence systems increasingly applied in consequential domains, researchers have begun to ask how AI systems ought to act in ethically charged situations where even humans lack consensus. In the Moral Machine project, researchers crowdsourced answers to "Trolley Problems" concerning autonomous vehicles. Subsequently, Noothigattu et al. (2018) proposed inferring linear functions that approximate each individual's preferences and aggregating these linear models by averaging parameters across the population. In this paper, we examine this averaging mechanism, focusing on fairness concerns and strategic effects. We investigate a simple setting where the population consists of two groups, the minority constitutes an α < 0.5 share of the population, and within-group preferences are homogeneous. Focusing on the fraction of contested cases where the minority group prevails, we make the following observations: (a) even when all parties report their preferences truthfully, the fraction of disputes where the minority prevails is less than proportionate in α; (b) the degree of sub-proportionality grows more severe as the level of disagreement between the groups increases; (c) when parties report preferences strategically, pure strategy equilibria do not always exist; and (d) whenever a pure strategy equilibrium exists, the majority group prevails 100% of the time. These findings raise concerns about stability and fairness of averaging as a mechanism for aggregating diverging voices. Finally, we discuss alternatives, including randomized dictatorship and median-based mechanisms.

For most natural language processing tasks, the dominant practice is to finetune large pretrained transformer models (e.g., BERT) using smaller downstream datasets.Despite the success of this approach, it remains unclear to what extent these gains are attributable to the massive background corpora employed for pretraining versus to the pretraining objectives themselves.This paper introduces a large-scale study of self-pretraining, where the same (downstream) training data is used for both pretraining and finetuning.In experiments addressing both ELECTRA and RoBERTa models and 10 distinct downstream classification datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around 10×-500× less data), outperforming the latter on 7 and 5 datasets, respectively.Surprisingly, these task-specific pretrained models often perform well on other tasks, including the GLUE benchmark.Self-pretraining also provides benefits on structured output prediction tasks such as question answering and commonsense inference, often providing more than 50% improvements compared to standard pretraining.Our results hint that often performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the use of external pretraining data in massive amounts.These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data. 1

The AI Ethics community faces an imperative to empower stakeholders and impacted community members so that they can scrutinize and influence the design, development, and use of AI systems in high-stakes domains. While a growing chorus of recent papers has kindled interest in so-called “participatory ML” methods, precisely what form participation ought to take and how to operationalize these ambitions are seldom addressed. Our survey of the relevant literature shows that in many papers, participation is reduced to highly structured, computational mechanisms designed to elicit mathematically tractable approximations of narrowly-defined moral values. Of papers that actually engage with real people, these engagements typically consist of one-time interactions with individuals that are often unrepresentative of the relevant stakeholders. Motivated by these clear limitations, we introduce a consolidated set of axes to evaluate and improve participatory approaches. We use these axes to analyze contemporary work in this space and outline future AI research directions that could meaningfully contribute to operationalizing the ideal of participation.

Rates of missing data often depend on record-keeping policies and thus may change across times and locations, even when the underlying features are comparatively stable. In this paper, we introduce the problem of Domain Adaptation under Missingness Shift (DAMS). Here, (labeled) source data and (unlabeled) target data would be exchangeable but for different missing data mechanisms. We show that if missing data indicators are available, DAMS reduces to covariate shift. Addressing cases where such indicators are absent, we establish the following theoretical results for underreporting completely at random: (i) covariate shift is violated (adaptation is required); (ii) the optimal linear source predictor can perform arbitrarily worse on the target domain than always predicting the mean; (iii) the optimal target predictor can be identified, even when the missingness rates themselves are not; and (iv) for linear models, a simple analytic adjustment yields consistent estimates of the optimal target parameters. In experiments on synthetic and semi-synthetic data, we demonstrate the promise of our methods when assumptions hold. Finally, we discuss a rich family of future extensions.

Personalized recommender systems suffuse modern life, shaping what media we read and what products we consume. Algorithms powering such systems tend to consist of supervised-learning-based heuristics, such as latent factor models with a variety of heuristically chosen prediction targets. Meanwhile, theoretical treatments of recommendation frequently address the decision-theoretic nature of the problem, including the need to balance exploration and exploitation, via the multi-armed bandits (MABs) framework. However, MAB-based approaches rely heavily on assumptions about human preferences. These preference assumptions are seldom tested using human subject studies, partly due to the lack of publicly available toolkits to conduct such studies. In this work, we conduct a study with crowdworkers in a comics recommendation MABs setting. Each arm represents a comic category, and users provide feedback after each recommendation. We check the validity of core MABs assumptions—that human preferences (reward distributions) are fixed over time—and find that they do not hold. This finding suggests that any MAB algorithm used for recommender systems should account for human preference dynamics. While answering these questions, we provide a flexible experimental framework for understanding human preference dynamics and testing MABs algorithms with human users. The code for our experimental framework and the collected data can be found at https://github.com/HumainLab/human-bandit-evaluation.

In recent years, NLP practitioners have converged on the following practice: (i) import an off-the-shelf pretrained (masked) language model; (ii) append a multilayer perceptron atop the CLS token's hidden representation (with randomly initialized weights); and (iii) finetune the entire model on a downstream task (MLP-FT).This procedure has produced massive gains on standard NLP benchmarks, but these models remain brittle, even to mild adversarial perturbations.In this work, we demonstrate surprising gains in adversarial robustness enjoyed by Model-tuning Via Prompts (MVP), an alternative method of adapting to downstream tasks.Rather than appending an MLP head to make output prediction, MVP appends a prompt template to the input, and makes prediction via text infilling/completion. Across 5 NLP datasets, 4 adversarial attacks, and 3 different models, MVP improves performance against adversarial substitutions by an average of 8% over standard methods and even outperforms adversarial training-based state-of-art defenses by 3.5%.By combining MVP with adversarial training, we achieve further improvements in adversarial robustness while maintaining performance on unperturbed examples.Finally, we conduct ablations to investigate the mechanism underlying these gains.Notably, we find that the main causes of vulnerability of MLP-FT can be attributed to the misalignment between pre-training and fine-tuning tasks, and the randomly initialized MLP parameters. 1

Despite the emergence of principled methods for domain adaptation under label shift, their sensitivity to shifts in class conditional distributions is precariously under explored. Meanwhile, popular deep domain adaptation heuristics tend to falter when faced with label proportions shifts. While several papers modify these heuristics in attempts to handle label proportions shifts, inconsistencies in evaluation standards, datasets, and baselines make it difficult to gauge the current best practices. In this paper, we introduce RLSbench, a large-scale benchmark for *relaxed label shift*, consisting of $>$500 distribution shift pairs spanning vision, tabular, and language modalities, with varying label proportions. Unlike existing benchmarks, which primarily focus on shifts in class-conditional $p(x|y)$, our benchmark also focuses on label marginal shifts. First, we assess 13 popular domain adaptation methods, demonstrating more widespread failures under label proportion shifts than were previously known. Next, we develop an effective two-step meta-algorithm that is compatible with most domain adaptation heuristics: (i) *pseudo-balance* the data at each epoch; and (ii) adjust the final classifier with target label distribution estimate. The meta-algorithm improves existing domain adaptation heuristics under large label proportion shifts, often by 2--10% accuracy points, while conferring minimal effect ($<$0.5%) when label proportions do not shift. We hope that these findings and the availability of RLSbench will encourage researchers to rigorously evaluate proposed methods in relaxed label shift settings. Code is publicly available at https://github.com/acmi-lab/RLSbench.

Recent efforts at explaining the interplay of memorization and generalization in deep overparametrized networks have posited that neural networks *memorize* ``hard'' examples in the final few layers of the model. Memorization refers to the ability to correctly predict on *atypical* examples of the training set. In this work, we show that rather than being confined to individual layers, memorization is a phenomenon confined to a small set of neurons in various layers of the model. First, via three experimental sources of converging evidence, we find that most layers are redundant for the memorization of examples and the layers that contribute to example memorization are, in general, not the final layers. The three sources are *gradient accounting* (measuring the contribution to the gradient norms from memorized and clean examples), *layer rewinding* (replacing specific model weights of a converged model with previous training checkpoints), and *retraining* (training rewound layers only on clean examples). Second, we ask a more generic question: can memorization be localized *anywhere* in a model? We discover that memorization is often confined to a small number of neurons or channels (around 5) of the model. Based on these insights we propose a new form of dropout---*example-tied dropout* that enables us to direct the memorization of examples to an aprior determined set of neurons. By dropping out these neurons, we are able to reduce the accuracy on memorized examples from 100% to 3%, while also reducing the generalization gap.

A burgeoning line of research has developed deep neural networks capable of approximating the solutions to high dimensional PDEs, opening related lines of theoretical inquiry focused on explaining how it is that these models appear to evade the curse of dimensionality. However, most theoretical analyses thus far have been limited to linear PDEs. In this work, we take a step towards studying the representational power of neural networks for approximating solutions to nonlinear PDEs. We focus on a class of PDEs known as *nonlinear elliptic variational PDEs*, whose solutions minimize an *Euler-Lagrange* energy functional $\mathcal{E}(u) = \int_\Omega L(x, u(x), \nabla u(x)) - f(x) u(x)dx$. We show that if composing a function with Barron norm $b$ with partial derivatives of $L$ produces a function of Barron norm at most $B_L b^p$, the solution to the PDE can be $\epsilon$-approximated in the $L^2$ sense by a function with Barron norm $O\left(\left(dB_L\right)^{\max\{p \log(1/ \epsilon), p^{\log(1/\epsilon)}\}}\right)$. By a classical result due to Barron (1993), this correspondingly bounds the size of a 2-layer neural network needed to approximate the solution. Treating $p, \epsilon, B_L$ as constants, this quantity is polynomial in dimension, thus showing neural networks can evade the curse of dimensionality. Our proof technique involves neurally simulating (preconditioned) gradient in an appropriate Hilbert space, which converges exponentially fast to the solution of the PDE, and such that we can bound the increase of the Barron norm at each iterate. Our results subsume and substantially generalize analogous prior results for linear elliptic PDEs over a unit hypercube.

Open vocabulary models (e.g. CLIP) have shown strong performance on zero-shot classification through their ability generate embeddings for each class based on their (natural language) names. Prior work has focused on improving the accuracy of these models through prompt engineering or by incorporating a small amount of labeled downstream data (via finetuning). However, there has been little focus on improving the richness of the class names themselves, which can pose issues when class labels are coarsely-defined and are uninformative. We propose Classification with Hierarchical Label Sets (or CHiLS), an alternative strategy for zero-shot classification specifically designed for datasets with implicit semantic hierarchies. CHiLS proceeds in three steps: (i) for each class, produce a set of subclasses, using either existing label hierarchies or by querying GPT-3; (ii) perform the standard zero-shot CLIP procedure as though these subclasses were the labels of interest; (iii) map the predicted subclass back to its parent to produce the final prediction. Across numerous datasets with underlying hierarchical structure, CHiLS leads to improved accuracy in situations both with and without ground-truth hierarchical information. CHiLS is simple to implement within existing zero-shot pipelines and requires no additional training cost. Code is available at: https://github.com/acmi-lab/CHILS.

Jacob Tyo, Bhuwan Dhingra, Zachary C. Lipton. Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

This paper focuses on supervised and unsupervised online label shift, where the class marginals $Q(y)$ varies but the class-conditionals $Q(x|y)$ remain invariant. In the unsupervised setting, our goal is to adapt a learner, trained on some offline labeled data, to changing label distributions given unlabeled online data. In the supervised setting, we must both learn a classifier and adapt to the dynamically evolving class marginals given only labeled online data. We develop novel algorithms that reduce the adaptation problem to online regression and guarantee optimal dynamic regret without any prior knowledge of the extent of drift in the label distribution. Our solution is based on bootstrapping the estimates of *online regression oracles* that track the drifting proportions. Experiments across numerous simulated and real-world online label shift scenarios demonstrate the superior performance of our proposed approaches, often achieving 1-3% improvement in accuracy while being sample and computationally efficient. Code is publicly available at https://github.com/Anon-djiwh/OnlineLabelShift

Self-training and contrastive learning have emerged as leading techniques for incorporating unlabeled data, both under distribution shift (unsupervised domain adaptation) and when it is absent (semi-supervised learning). However, despite the popularity and compatibility of these techniques, their efficacy in combination remains surprisingly unexplored. In this paper, we first undertake a systematic empirical investigation of this combination, finding (i) that in domain adaptation settings, self-training and contrastive learning offer significant complementary gains; and (ii) that in semi-supervised learning settings, surprisingly, the benefits are not synergistic. Across eight distribution shift datasets (e.g., BREEDs, WILDS), we demonstrate that the combined method obtains 3--8\% higher accuracy than either approach independently. Finally, we theoretically analyze these techniques in a simplified model of distribution shift demonstrating scenarios under which the features produced by contrastive learning can yield a good initialization for self-training to further amplify gains and achieve optimal performance, even when either method alone would fail.

Data-driven machine learning approaches are being increasingly used to solve partial differential equations (PDEs). They have shown particularly striking successes when training an operator, which takes as input a PDE in some family, and outputs its solution. However, the architectural design space, especially given structural knowledge of the PDE family of interest, is still poorly understood. We seek to remedy this gap by studying the benefits of weight-tied neural network architectures for steady-state PDEs. To achieve this, we first demonstrate that the solution of most steady-state PDEs can be expressed as a fixed point of a non-linear operator. Motivated by this observation, we propose FNO-DEQ, a deep equilibrium variant of the FNO architecture that directly solves for the solution of a steady-state PDE as the infinite-depth fixed point of an implicit operator layer using a black-box root solver and differentiates analytically through this fixed point resulting in $\mathcal{O}(1)$ training memory. Our experiments indicate that FNO-DEQ-based architectures outperform FNO-based baselines with $4\times$ the number of parameters in predicting the solution to steady-state PDEs such as Darcy Flow and steady-state incompressible Navier-Stokes. Finally, we show FNO-DEQ is more robust when trained with datasets with more noisy observations than the FNO-based baselines, demonstrating the benefits of using appropriate inductive biases in architectural design for different neural network based PDE solvers. Further, we show a universal approximation result that demonstrates that FNO-DEQ can approximate the solution to any steady-state PDE that can be written as a fixed point equation.

We introduce the notion of risk-limiting financial audits (RLFA): procedures that manually evaluate a subset of $N$ financial transactions to check the validity of a claimed assertion $\mathcal{A}$ about the transactions. More specifically, RLFA satisfy two properties: (i) if $\mathcal{A}$ is false, they correctly disprove it with probability at least $1-\delta$, and (ii) they validate the correctness of $\mathcal{A}$ with probability $1$, if it is true. We propose a general RLFA strategy, by constructing new confidence sequences (CSs) for the weighted average of $N$ unknown values, based on samples drawn without replacement from a (randomized) weighted sampling scheme. Next, we develop methods to improve the quality of CSs by incorporating side information about the unknown values. We show that when the side information is sufficiently accurate, it can directly drive the sampling. For the case where the accuracy is unknown a priori, we introduce an alternative approach using control variates. Crucially, our construction adapts to the quality of side information by strongly leveraging the side information if it is highly predictive, and learning to ignore it if it is uninformative. Our methods also recover the state-of-the-art bounds for the special case of uniformly sampled observations with no side information, which has already found applications in election auditing. The harder weighted case with general side information solves the more challenging problem of AI-assisted financial auditing.

In attempts to "explain" predictions of machine learning models, researchers have proposed hundreds of techniques for attributing predictions to features that are deemed important. While these attributions are often claimed to hold the potential to improve human "understanding" of the models, surprisingly little work explicitly evaluates progress towards this aspiration. In this paper, we conduct a crowdsourcing study, where participants interact with deception detection models that have been trained to distinguish between genuine and fake hotel reviews. They are challenged both to simulate the model on fresh reviews, and to edit reviews with the goal of lowering the probability of the originally predicted class. Successful manipulations would lead to an adversarial example. During the training (but not the test) phase, input spans are highlighted to communicate salience. Through our evaluation, we observe that for a linear bag-of-words model, participants with access to the feature coefficients during training are able to cause a larger reduction in model confidence in the testing phase when compared to the no-explanation control. For the BERT-based classifier, popular local explanations do not improve their ability to reduce the model confidence over the no-explanation case. Remarkably, when the explanation for the BERT model is given by the (global) attributions of a linear model trained to imitate the BERT model, people can effectively manipulate the model.

Traditionally, when recommender systems are formalized as multi-armed bandits, the policy of the recommender system influences the rewards accrued, but not the length of interaction. However, in real-world systems, dissatisfied users may depart (and never come back). In this work, we propose a novel multi-armed bandit setup that captures such policy-dependent horizons. Our setup consists of a finite set of user types, and multiple arms with Bernoulli payoffs. Each (user type, arm) tuple corresponds to an (unknown) reward probability. Each user's type is initially unknown and can only be inferred through their response to recommendations. Moreover, if a user is dissatisfied with their recommendation, they might depart the system. We first address the case where all users share the same type, demonstrating that a recent UCB-based algorithm is optimal. We then move forward to the more challenging case, where users are divided among two types. While naive approaches cannot handle this setting, we provide an efficient learning algorithm that achieves O(sqrt(T)ln(T)) regret, where T is the number of users.

Addressing such diverse ends as mitigating safety risks, aligning agent behavior with human preferences, and improving the efficiency of learning, an emerging line of reinforcement learning research addresses the entire distribution of returns and various risk functionals that depend upon it. In the contextual bandit setting, recently work on off-policy risk assessment estimates the target policy’s CDF of returns, providing finite sample guarantees that extend to (and hold simultaneously over) plugin estimates of an arbitrarily large set of risk functionals. In this paper, we lift OPRA to Markov decision processes (MDPs), where importance sampling (IS) CDF estimators suffer high variance on longer trajectories due to vanishing (and exploding) importance weights. To mitigate these problems, we incorporate model-based estimation to develop the first doubly robust (DR) estimator for the CDF of returns in MDPs. The DR estimator enjoys significantly less variance and, when the model is well specified, achieves the Cramer-Rao variance lower bound. Moreover, for many risk functionals, the downstream estimates enjoy both lower bias and lower variance. Additionally, we derive the first minimax lower bounds for off-policy CDF and risk estimation, which match our error bounds up to a constant. Finally, we demonstrate the efficacy of our DR CDF estimates experimentally on several different environments.

Standard uniform convergence results bound the generalization gap of the expected loss over a hypothesis class. The emergence of risk-sensitive learning requires generalization guarantees for functionals of the loss distribution beyond the expectation. While prior works specialize in uniform convergence of particular functionals, our work provides uniform convergence for a general class of Hölder risk functionals for which the closeness in the Cumulative Distribution Function (CDF) entails closeness in risk. We establish the first uniform convergence results for estimating the CDF of the loss distribution, which yield uniform convergence guarantees that hold simultaneously both over a class of Hölder risk functionals and over a hypothesis class. Thus licensed to perform empirical risk minimization, we develop practical gradient-based methods for minimizing distortion risks (widely studied subset of Hölder risks that subsumes the spectral risks, including the mean, conditional value at risk, cumulative prospect theory risks, and others) and provide convergence guarantees. In experiments, we demonstrate the efficacy of our learning procedure, both in settings where uniform convergence results hold and in high-dimensional settings with deep networks.

We introduce the problem of domain adaptation under Open Set Label Shift (OSLS), where the label distribution can change arbitrarily and a new class may arrive during deployment, but the class-conditional distributions $p(x|y)$ are domain-invariant. OSLS subsumes domain adaptation under label shift and Positive-Unlabeled (PU) learning. The learner's goals here are two-fold: (a) estimate the target label distribution, including the novel class; and (b) learn a target classifier. First, we establish the necessary and sufficient for identifying these quantities. Second, motivated by advances in label shift and PU learning, we propose practical methods for both tasks that leverage black-box predictors. Unlike typical Open Set Domain Adaptation (OSDA) problems, which tend to be ill-posed and amenable only to heuristics, OSLS offers a well-posed problem amenable to more principled machinery. Experiments across numerous semi-synthetic benchmarks on vision, language, and medical datasets demonstrate that our methods consistently outperform OSDA baselines, achieving $10$--$25\%$ improvements in target domain accuracy. Finally, we analyze the proposed methods, establishing finite-sample convergence to the true label marginal and convergence to optimal classifier for linear models in a Gaussian setup. Code is available at https://github.com/acmi-lab/Open-Set-Label-Shift.

Researchers investigating example hardness have increasingly focused on the dynamics by which neural networks learn and forget examples throughout training. Popular metrics derived from these dynamics include (i) the epoch at which examples are first correctly classified; (ii) the number of times their predictions flip during training; and (iii) whether their prediction flips if they are held out. However, these metrics do not distinguish among examples that are hard for distinct reasons, such as membership in a rare subpopulation, being mislabeled, or belonging to a complex subpopulation. In this paper, we propose second-split forgetting time (SSFT), a complementary metric that tracks the epoch (if any) after which an original training example is forgotten as the network is fine-tuned on a randomly held out partition of the data. Across multiple benchmark datasets and modalities, we demonstrate that mislabeled examples are forgotten quickly, and seemingly rare examples are forgotten comparatively slowly. By contrast, metrics only considering the first split learning dynamics struggle to differentiate the two. At large learning rates, SSFT tends to be robust across architectures, optimizers, and random seeds. From a practical standpoint, the SSFT can (i) help to identify mislabeled samples, the removal of which improves generalization; and (ii) provide insights about failure modes. Through theoretical analysis addressing overparameterized linear models, we provide insights into how the observed phenomena may arise.

What sorts of structure might enable a learner to discover classes from unlabeled data? Traditional approaches rely on feature-space similarity and heroic assumptions on the data. In this paper, we introduce unsupervised learning under Latent Label Shift (LLS), where the label marginals $p_d(y)$ shift but the class conditionals $p(x|y)$ do not. This work instantiates a new principle for identifying classes: elements that shift together group together. For finite input spaces, we establish an isomorphism between LLS and topic modeling: inputs correspond to words, domains to documents, and labels to topics. Addressing continuous data, we prove that when each label's support contains a separable region, analogous to an anchor word, oracle access to $p(d|x)$ suffices to identify $p_d(y)$ and $p_d(y|x)$ up to permutation. Thus motivated, we introduce a practical algorithm that leverages domain-discriminative models as follows: (i) push examples through domain discriminator $p(d|x)$; (ii) discretize the data by clustering examples in $p(d|x)$ space; (iii) perform non-negative matrix factorization on the discrete data; (iv) combine the recovered $p(y|d)$ with the discriminator outputs $p(d|x)$ to compute $p_d(y|x) \; \forall d$. With semisynthetic experiments, we show that our algorithm can leverage domain information to improve upon competitiveunsupervised classification methods. We reveal a failure mode of standard unsupervised classification methods when data-space similarity does not indicate true groupings, and show empirically that our method better handles this case. Our results establish a deep connection between distribution shift and topic modeling, opening promising lines for future work.

Abstract While many methods purport to explain predictions by highlighting salient features, what aims these explanations serve and how they ought to be evaluated often go unstated. In this work, we introduce a framework to quantify the value of explanations via the accuracy gains that they confer on a student model trained to simulate a teacher model. Crucially, the explanations are available to the student during training, but are not available at test time. Compared with prior proposals, our approach is less easily gamed, enabling principled, automatic, model-agnostic evaluation of attributions. Using our framework, we compare numerous attribution methods for text classification and question answering, and observe quantitative differences that are consistent (to a moderate to high degree) across different student model architectures and learning strategies.1

Autoregressive models using Transformers have emerged as the dominant approach for music generation with the goal of synthesizing minute-long compositions that exhibit large-scale musical structure. These models are commonly trained by minimizing the negative log-likelihood (NLL) of the observed sequence in an autoregressive manner. Unfortunately, the quality of samples from these models tends to degrade significantly for long sequences, a phenomenon attributed to exposure bias. Fortunately, we are able to detect these failures with classifiers trained to distinguish between real and sampled sequences, an observation that motivates our exploration of adversarial losses to complement the NLL objective. We use a pre-trained Span-BERT model for the discriminator of the GAN, which in our experiments helped with training stability. We use the Gumbel-Softmax trick to obtain a differentiable approximation of the sampling process. This makes discrete sequences amenable to optimization in GANs. In addition, we break the sequences into smaller chunks to ensure that we stay within a given memory budget. We demonstrate via human evaluations and a new discriminative metric that the music generated by our approach outperforms a baseline trained with likelihood maximization, the state-of-the-art Music Transformer, and other GANs used for sequence generation. 57% of people prefer music generated via our approach while 43% prefer Music Transformer.

Divyansh Kaushik, Douwe Kiela, Zachary C. Lipton, Wen-tau Yih. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Kundan Krishna, Sopan Khosla, Jeffrey Bigham, Zachary C. Lipton. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Typically, fair machine learning research focuses on a single decision maker and assumes that the underlying population is stationary. However, many of the critical domains motivating this work are characterized by competitive marketplaces with many decision makers. Realistically, we might expect only a subset of them to adopt any non-compulsory fairness-conscious policy, a situation that political philosophers call partial compliance. This possibility raises important questions: how does partial compliance and the consequent strategic behavior of decision subjects affect the allocation outcomes? If k% of employers were to voluntarily adopt a fairness-promoting intervention, should we expect k% progress (in aggregate) towards the benefits of universal adoption, or will the dynamics of partial compliance wash out the hoped-for benefits? How might adopting a global (versus local) perspective impact the conclusions of an auditor? In this paper, we propose a simple model of an employment market, leveraging simulation as a tool to explore the impact of both interaction effects and incentive effects on outcomes and auditing metrics. Our key findings are that at equilibrium: (1) partial compliance by k% of employers can result in far less than proportional (k%) progress towards the full compliance outcomes; (2) the gap is more severe when fair employers match global (vs local) statistics; (3) choices of local vs global statistics can paint dramatically different pictures of the performance vis-a-vis fairness desiderata of compliant versus non-compliant employers; (4) partial compliance based on local parity measures can induce extreme segregation. Finally, we discuss implications for auditors and insights concerning the design of regulatory frameworks.

Re-offense risk is considered in decision-making at many stages of the criminal justice system, from pre-trial, to sentencing, to parole. To aid decision-makers in their assessments, institutions increasingly rely on algorithmic risk assessment instruments (RAIs). These tools assess the likelihood that an individual will be arrested for a new criminal offense within some time window following their release. However, since not all crimes result in arrest, RAIs do not directly assess the risk of re-offense. Furthermore, disparities in the likelihood of arrest can potentially lead to biases in the resulting risk scores. Several recent validations of RAIs have therefore focused on arrests for violent offenses, which are viewed as being more accurate and less biased reflections of offending behavior. In this paper, we investigate biases in violent arrest data by analysing racial disparities in the likelihood of arrest for White and Black violent offenders. We focus our study on 2007--2016 incident-level data of violent offenses from 16 US states as recorded in the National Incident Based Reporting System (NIBRS). Our analysis shows that the magnitude and direction of the racial disparities depend on various characteristics of the crimes. In addition, our investigation reveals large variations in arrest rates across geographical locations and offense types. We discuss the implications of the observed disconnect between re-arrest and re-offense in the context of RAIs and the challenges around the use of data from NIBRS to correct for the sampling bias.

Given only data generated by a standard confounding graph with unobserved confounder, the Average Treatment Effect (ATE) is not identifiable. To estimate the ATE, a practitioner must then either (a) collect deconfounded data; (b) run a clinical trial; or (c) elucidate further properties of the causal graph that might render the ATE identifiable. In this paper, we consider the benefit of incorporating a large confounded observational dataset (confounder unobserved) alongside a small deconfounded observational dataset (confounder revealed) when estimating the ATE. Our theoretical results suggest that the inclusion of confounded data can significantly reduce the quantity of deconfounded data required to estimate the ATE to within a desired accuracy level. Moreover, in some cases—say, genetics—we could imagine retrospectively selecting samples to deconfound. We demonstrate that by actively selecting these samples based upon the (already observed) treatment and outcome, we can reduce sample complexity further. Our theoretical and empirical results establish that the worst-case relative performance of our approach (vs. a natural benchmark) is bounded while our best-case gains are unbounded. Finally, we demonstrate the benefits of selective deconfounding using a large real-world dataset related to genetic mutation in cancer.

Unsupervised Data Augmentation (UDA) is a semi-supervised technique that applies a consistency loss to penalize differences between a model's predictions on (a) observed (unlabeled) examples; and (b) corresponding 'noised' examples produced via data augmentation.While UDA has gained popularity for text classification, open questions linger over which of its components are important, and how to extend the method to sequence labeling tasks; this paper addresses these questions.Our main contribution is an empirical study of UDA to establish which components of the algorithm confer benefits in NLP.Notably, although prior work has emphasized use of clever augmentation techniques including back-translation, we find that enforcing consistency between predictions assigned to observed and randomly substituted words often yields comparable (or greater) benefits compared to these more complex perturbation models.Furthermore, we find that applying UDA's consistency loss affords meaningful gains without any unlabeled data at all, i.e., in a standard supervised setting.In short, UDA need not be unsupervised to realize much of its noted benefits, and does not require complex data augmentation to be effective.

To assess generalization, machine learning scientists typically either (i) bound the generalization gap and then (after training) plug in the empirical risk to obtain a bound on the true risk; or (ii) validate empirically on holdout data. However, (i) typically yields vacuous guarantees for overparameterized models; and (ii) shrinks the training set and its guarantee erodes with each re-use of the holdout set. In this paper, we leverage unlabeled data to produce generalization bounds. After augmenting our (labeled) training set with randomly labeled data, we train in the standard fashion. Whenever classifiers achieve low error on the clean data but high error on the random data, our bound ensures that the true risk is low. We prove that our bound is valid for 0-1 empirical risk minimization and with linear classifiers trained by gradient descent. Our approach is especially useful in conjunction with deep learning due to the early learning phenomenon whereby networks fit true labels before noisy labels but requires one intuitive assumption. Empirically, on canonical computer vision and NLP tasks, our bound provides non-vacuous generalization guarantees that track actual performance closely. This work enables practitioners to certify generalization even when (labeled) holdout data is unavailable and provides insights into the relationship between random label noise and generalization.

Modern policy gradient algorithms such as Proximal Policy Optimization (PPO) rely on an arsenal of heuristics, including loss clipping and gradient clipping, to ensure successful learning. These heuristics are reminiscent of techniques from robust statistics, commonly used for estimation in outlier-rich ("heavy-tailed") regimes. In this paper, we present a detailed empirical study to characterize the heavy-tailed nature of the gradients of the PPO surrogate reward function. We demonstrate that the gradients, especially for the actor network, exhibit pronounced heavy-tailedness and that it increases as the agent’s policy diverges from the behavioral policy (i.e., as the agent goes further off policy). Further examination implicates the likelihood ratios and advantages in the surrogate reward as the main sources of the observed heavy-tailedness. We then highlight issues arising due to the heavy-tailed nature of the gradients. In this light, we study the effects of the standard PPO clipping heuristics, demonstrating that these tricks primarily serve to offset heavy-tailedness in gradients. Thus motivated, we propose incorporating GMOM, a high-dimensional robust estimator, into PPO as a substitute for three clipping tricks. Despite requiring less hyperparameter tuning, our method matches the performance of PPO (with all heuristics enabled) on a battery of MuJoCo continuous control tasks.

Link prediction methods are frequently applied in recommender systems, e.g., to suggest citations for academic papers or friends in social networks. However, exposure bias can arise when users are systematically underexposed to certain relevant items. For example, in citation networks, authors might be more likely to encounter papers from their own field and thus cite them preferentially. This bias can propagate through naively trained link predictors, leading to both biased evaluation and high generalization error (as assessed by true relevance). Moreover, this bias can be exacerbated by feedback loops. We propose estimators that leverage known exposure probabilities to mitigate this bias and consequent feedback loops. Next, we provide a loss function for learning the exposure probabilities from data. Finally, experiments on semi-synthetic data based on real-world citation networks, show that our methods reliably identify (truly) relevant citations. Additionally, our methods lead to greater diversity in the recommended papers’ fields of study. The code is available at github.com/shantanu95/exposure-bias-link-rec.

Given only positive examples and unlabeled examples (from both positive and negative classes), we might hope nevertheless to estimate an accurate positive-versus-negative classifier. Formally, this task is broken down into two subtasks: (i) Mixture Proportion Estimation (MPE)---determining the fraction of positive examples in the unlabeled data; and (ii) PU-learning---given such an estimate, learning the desired positive-versus-negative classifier. Unfortunately, classical methods for both problems break down in high-dimensional settings. Meanwhile, recently proposed heuristics lack theoretical coherence and depend precariously on hyperparameter tuning. In this paper, we propose two simple techniques: Best Bin Estimation (BBE) (for MPE); and Conditional Value Ignoring Risk (CVIR), a simple objective for PU-learning. Both methods dominate previous approaches empirically, and for BBE, we establish formal guarantees that hold whenever we can train a model to cleanly separate out a small subset of positive examples. Our final algorithm (TED)$^n$, alternates between the two procedures, significantly improving both our mixture proportion estimator and classifier

Researchers often face data fusion problems, where multiple data sources are available, each capturing a distinct subset of variables. While problem formulations typically take the data as given, in practice, data acquisition can be an ongoing process. In this paper, we introduce the problem of deciding, at each time, which data source to sample from. Our goal is to estimate a given functional of the parameters of a probabilistic model as efficiently as possible. We propose online moment selection (OMS), a framework in which structural assumptions are encoded as moment conditions. The optimal action at each step depends, in part, on the very moments that identify the functional of interest. Our algorithms balance exploration with choosing the best action as suggested by estimated moments. We propose two selection strategies: (1) explore-then-commit (ETC) and (2) explore-then-greedy (ETG), proving that both achieve zero asymptotic regret as assessed by MSE. We instantiate our setup for average treatment effect estimation, where structural assumptions are given by a causal graph and data sources include subsets of mediators, confounders, and instrumental variables.

Even when unable to run experiments, practitioners can evaluate prospective policies, using previously logged data. However, while the bandits literature has adopted a diverse set of objectives, most research on off-policy evaluation to date focuses on the expected reward. In this paper, we introduce Lipschitz risk functionals, a broad class of objectives that subsumes conditional value-at-risk (CVaR), variance, mean-variance, many distorted risks, and CPT risks, among others. We propose Off-Policy Risk Assessment (OPRA), a framework that first estimates a target policy's CDF and then generates plugin estimates for any collection of Lipschitz risks, providing finite sample guarantees that hold simultaneously over the entire class. We instantiate OPRA with both importance sampling and doubly robust estimators. Our primary theoretical contributions are (i) the first uniform concentration inequalities for both CDF estimators in contextual bandits and (ii) error bounds on our Lipschitz risk estimates, which all converge at a rate of $O(1/\sqrt{n})$.

Psychological research shows that enjoyment of many goods is subject to satiation, with short-term satisfaction declining after repeated exposures to the same item. Nevertheless, proposed algorithms for powering recommender systems seldom model these dynamics, instead proceeding as though user preferences were fixed in time. In this work, we introduce rebounding bandits, a multi-armed bandit setup, where satiation dynamics are modeled as time-invariant linear dynamical systems. Expected rewards for each arm decline monotonically with consecutive exposures and rebound towards the initial reward whenever that arm is not pulled. Unlike classical bandit algorithms, methods for tackling rebounding bandits must plan ahead and model-based methods rely on estimating the parameters of the satiation dynamics. We characterize the planning problem, showing that the greedy policy is optimal when the arms exhibit identical deterministic dynamics. To address stochastic satiation dynamics with unknown parameters, we propose Explore-Estimate-Plan, an algorithm that pulls arms methodically, estimates the system dynamics, and then plans accordingly.

Recent experiments have shown that deep networks can approximate solutions to high-dimensional PDEs, seemingly escaping the curse of dimensionality. However, questions regarding the theoretical basis for such approximations, including the required network size remain open. In this paper, we investigate the representational power of neural networks for approximating solutions to linear elliptic PDEs with Dirichlet boundary conditions. We prove that when a PDE's coefficients are representable by small neural networks, the parameters required to approximate its solution scale polynomially with the input dimension $d$ and proportionally to the parameter counts of the coefficient networks. To this end, we develop a proof technique that simulates gradient descent (in an appropriate Hilbert space) by growing a neural network architecture whose iterates each participate as sub-networks in their (slightly larger) successors, and converge to the solution of the PDE.

Given a causal graph, the do-calculus can express treatment effects as functionals of the observational joint distribution that can be estimated empirically. Sometimes the do-calculus identifies multiple valid formulae, prompting us to compare the statistical properties of the corresponding estimators. For example, the backdoor formula applies when all confounders are observed and the frontdoor formula applies when an observed mediator transmits the causal effect. In this paper, we investigate the over-identified scenario where both confounders and mediators are observed, rendering both estimators valid. Addressing the linear Gaussian causal model, we demonstrate that either estimator can dominate the other by an unbounded constant factor. Next, we derive an optimal estimator, which leverages all observed variables, and bound its finite-sample variance. We show that it strictly outperforms the backdoor and frontdoor estimators and that this improvement can be unbounded. We also present a procedure for combining two datasets, one with observed confounders and another with observed mediators. Finally, we evaluate our methods on both simulated data and the IHDP and JTPA datasets.

As algorithmic risk assessment instruments (RAIs) are increasingly adopted to assist decision makers, their predictive performance and potential to promote inequity have come under scrutiny. However, while most studies examine these tools in isolation, researchers have come to recognize that assessing their impact requires understanding the behavior of their human interactants. In this paper, building off of several recent crowdsourcing works focused on criminal justice, we conduct a vignette study in which laypersons are tasked with predicting future re-arrests. Our key findings are as follows: (1) Participants often predict that an offender will be rearrested even when they deem the likelihood of re-arrest to be well below 50%; (2) Participants do not anchor on the RAI's predictions; (3) The time spent on the survey varies widely across participants and most cases are assessed in less than 10 seconds; (4) Judicial decisions, unlike participants' predictions, depend in part on factors that are orthogonal to the likelihood of re-arrest. These results highlight the influence of several crucial but often overlooked design decisions and concerns around generalizability when constructing crowdsourcing studies to analyze the impacts of RAI

Attention mechanisms are ubiquitous components in neural architectures applied to natural language processing.In addition to yielding gains in predictive accuracy, attention weights are often claimed to confer interpretability, purportedly useful both for providing insights to practitioners and for explaining why a model makes its decisions to stakeholders.We call the latter use of attention mechanisms into question by demonstrating a simple method for training models to produce deceptive attention masks.Our method diminishes the total weight assigned to designated impermissible tokens, even when the models can be shown to nevertheless rely on these features to drive predictions.Across multiple models and tasks, our approach manipulates attention weights while paying surprisingly little cost in accuracy.Through a human study, we show that our manipulated attention-based explanations deceive people into thinking that predictions from a model biased against gender minorities do not rely on the gender.Consequently, our results cast doubt on attention's reliability as a tool for auditing algorithms in the context of fairness and accountability.1

Inspired by recent breakthroughs in predictive modeling, practitioners in both industry and government have turned to machine learning with hopes of operationalizing predictions to drive automated decisions. Unfortunately, many social desiderata concerning consequential decisions, such as justice or fairness, have no natural formulation within a purely predictive framework. In the hopes of mitigating these problems, researchers have proposed a variety of metrics for quantifying deviations from various statistical parities that we might hope to observe in a fair world, offering a variety of algorithms that attempt to satisfy subsets of these parities or to trade off the degree to which they are satisfied against utility. In this paper, we connect this approach to fair machine learning to the literature on ideal and non-ideal methodological approaches in political philosophy. The ideal approach requires positing the principles according to which a just world would operate. In the most straightforward application of ideal theory, one supports a proposed policy by arguing that it closes a discrepancy between the real and ideal worlds. However, by failing to account for the mechanisms by which our non-ideal world arose, the responsibilities of various decision-makers, and the impacts of their actions, naive applications of ideal thinking can lead to misguided policies. In this paper, we demonstrate a connection between the recent literature on fair machine learning and the ideal approach in political philosophy, and show that some recently uncovered shortcomings in proposed algorithms reflect broader troubles faced by the ideal approach. We work this analysis through for different formulations of fairness and conclude with a critical discussion of real-world impacts and directions for new research.

Modern multilingual models are trained on concatenated text from multiple languages in hopes of conferring benefits to each (positive transfer), with the most pronounced benefits accruing to low-resource languages.However, recent work has shown that this approach can degrade performance on high-resource languages, a phenomenon known as negative interference.In this paper, we present the first systematic study of negative interference.We show that, contrary to previous belief, negative interference also impacts low-resource languages.While parameters are maximally shared to learn language-universal structures, we demonstrate that language-specific parameters do exist in multilingual models and they are a potential cause of negative interference.Motivated by these observations, we also present a meta-learning algorithm that obtains better cross-lingual transferability and alleviates negative interference, by adding languagespecific layers as meta-parameters and training them in a manner that explicitly improves shared layers' generalization on all languages.Overall, our results show that negative interference is more common than previously known, suggesting new directions for improving multilingual representations. 1 Model NER (F1) POS (F1) ar fr ru hi sw te avg

On a periodic basis, publicly traded companies report fundamentals, financial data including revenue, earnings, debt, among others. Quantitative finance research has identified several factors, functions of the reported data that historically correlate with stock market performance. In this paper, we first show through simulation that if we could select stocks via factors calculated on future fundamentals (via oracle), that our portfolios would far outperform standard factor models. Motivated by this insight, we train deep nets to forecast future fundamentals from a trailing 5-year history. We propose lookahead factor models which plug these predicted future fundamentals into traditional factors. Finally, we incorporate uncertainty estimates from both neural heteroscedastic regression and a dropout-based heuristic, improving performance by adjusting our portfolios to avert risk. In retrospective analysis, we leverage an industry-grade portfolio simulator (backtester) to show simultaneous improvement in annualized return and Sharpe ratio. Specifically, the simulated annualized return for the uncertainty-aware model is 17.7% (vs 14.0% for a standard factor model) and the Sharpe ratio is 0.84 (vs 0.52).

Under label shift, the label distribution $p(y)$ might change but the class-conditional distributions $p(x|y)$ do not. There are two dominant approaches for estimating the label marginal. BBSE, a moment-matching approach based on confusion matrices, is provably consistent and provides interpretable error bounds. However, a maximum likelihood estimation approach, which we call MLLS, dominates empirically. In this paper, we present a unified view of the two methods and the first theoretical characterization of MLLS. Our contributions include (i) consistency conditions for MLLS, which include calibration of the classifier and a confusion matrix invertibility condition that BBSE also requires; (ii) a unified framework, casting BBSE as roughly equivalent to MLLS for a particular choice of calibration method; and (iii) a decomposition of MLLS's finite-sample error into terms reflecting miscalibration and estimation error. Our analysis attributes BBSE's statistical inefficiency to a loss of information due to coarse calibration. Experiments on synthetic data, MNIST, and CIFAR10 support our findings.

Questions of fairness and bias abound in all socially-consequential decisions pertaining to collection and management of data. Whether designing protocols for peer review of research papers, setting hiring policies, or framing research question in genetics, any data-management decision with the potential to allocate benefits or confer harms raises concerns about who gains or loses that may fail to surface in naively-chosen performance measures. Data science interacts with these questions in two fundamentally different ways: (i) as the technology driving the very systems responsible for certain social impacts, posing new questions about what it means for such systems to accord with ethical norms and the law; and (ii) as a set of powerful tools for analyzing existing data management systems, e.g., for auditing existing systems for various biases. This tutorial will tackle both angles on the interaction between technology and society vis-a-vis concerns over fairness and bias, particularly focusing on the collection and management of data. Our presentation will cover a wide range of disciplinary perspectives with the first part focusing on the social impacts of technology and the formulations of fairness and bias defined via protected characteristics and the second part taking a deep into peer review and distributed human evaluations, to explore other forms of bias, such as that due to subjectivity, miscalibration, and dishonest behavior.

Convolutional neural networks typically consist of many convolutional layers followed by one or more fully connected layers. While convolutional layers map between high-order activation tensors, the fully connected layers operate on flattened activation vectors. Despite empirical success, this approach has notable drawbacks. Flattening followed by fully connected layers discards multilinear structure in the activations and requires many parameters. We address these problems by incorporating tensor algebraic operations that preserve multilinear structure at every layer. First, we introduce Tensor Contraction Layers (TCLs) that reduce the dimensionality of their input while preserving their multilinear structure using tensor contraction. Next, we introduce Tensor Regression Layers (TRLs), which express outputs through a low-rank multilinear mapping from a high-order activation tensor to an output tensor of arbitrary order. We learn the contraction and regression factors end-to-end, and produce accurate nets with fewer parameters. Additionally, our layers regularize networks by imposing low-rank constraints on the activations (TCL) and regression weights (TRL). Experiments on ImageNet show that, applied to VGG and ResNet architectures, TCLs and TRLs reduce the number of parameters compared to fully connected layers by more than 65% while maintaining or increasing accuracy. In addition to the space savings, our approach's ability to leverage topological structure can be crucial for structured data such as MRI. In particular, we demonstrate significant performance improvements over comparable architectures on three tasks associated with the UK Biobank dataset.

To combat adversarial spelling mistakes, we propose placing a word recognition model in front of the downstream classifier.Our word recognition models build upon the RNN semicharacter architecture, introducing several new backoff strategies for handling rare and unseen words.Trained to recognize words corrupted by random adds, drops, swaps, and keyboard mistakes, our method achieves 32% relative (and 3.3% absolute) error reduction over the vanilla semi-character model.Notably, our pipeline confers robustness on the downstream classifier, outperforming both adversarial training and off-the-shelf spell checkers.Against a BERT model fine-tuned for sentiment analysis, a single adversarially-chosen character attack lowers accuracy from 90.3% to 45.8%.Our defense restores accuracy to 75% 1 .Surprisingly, better word recognition does not always entail greater robustness.Our analysis reveals that robustness also depends upon a quantity that we denote the sensitivity.

Alankar Jain, Bhargavi Paranjape, Zachary C. Lipton. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

David Lowell, Zachary C. Lipton, Byron C. Wallace. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Importance-weighted risk minimization is a key ingredient in many machine learning algorithms for causal inference, domain adaptation, class imbalance, and off-policy reinforcement learning. While the effect of importance weighting is well-characterized for low-capacity misspecified models, little is known about how it impacts over-parameterized, deep neural networks. This work is inspired by recent theoretical results showing that on (linearly) separable data, deep linear networks optimized by SGD learn weight-agnostic solutions, prompting us to ask, for realistic deep networks, for which many practical datasets are separable, what is the effect of importance weighting? We present the surprising finding that while importance weighting impacts models early in training, its effect diminishes over successive epochs. Moreover, while L2 regularization and batch normalization (but not dropout), restore some of the impact of importance weighting, they express the effect via (seemingly) the wrong abstraction: why should practitioners tweak the L2 regularization, and by how much, to produce the correct weighting effect? Our experiments confirm these findings across a range of architectures and datasets.

Domain adaptation addresses the common situation in which the target distribution generating our test data differs from the source distribution generating our training data. While absent assumptions, domain adaptation is impossible, strict conditions, e.g. covariate or label shift, enable principled algorithms. Recently-proposed domain-adversarial approaches consist of aligning source and target encodings, an approach often motivated as minimizing two (of three) terms in a theoretical bound on target error. Unfortunately, this minimization can cause arbitrary increases in the third term, a problem guaranteed to arise under shifting label distributions. We propose asymmetrically-relaxed distribution alignment, a new approach that overcomes some limitations of standard domain-adversarial algorithms. Moreover, we characterize precise assumptions under which our algorithm is theoretically principled and demonstrate empirical benefits on both synthetic and real datasets.

Every day, thousands of customers post questions on Amazon product pages. After some time, if they are fortunate, a knowledgeable customer might answer their question. Observing that many questions can be answered based upon the available product reviews, we propose the task of review-based QA. Given a corpus of reviews and a question, the QA system synthesizes an answer. To this end, we introduce a new dataset and propose a method that combines informational retrieval techniques for selecting relevant reviews (given a question) and "reading comprehension" models for synthesizing an answer (given a question and review). Our dataset consists of 923k questions, 3.6M answers and 14M reviews across 156k products. Building on the well-known Amazon dataset, we additionally collect annotations marking each question as either answerable or unanswerable based on the available reviews. A deployed system could first classify a question as answerable before attempting to generate a provisional answer. Notably, unlike many popular QA datasets, here the questions, passages, and answers are extracted from real human interactions. We evaluate a number of models for answer generation and propose strong baselines, demonstrating the challenging nature of this new task.

We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.

Despite their renowned in-domain predictive power, convolutional neural networks are known to rely more on high-frequency patterns that humans deem superficial than on low-frequency patterns that agree better with intuitions about what constitutes category membership. This paper proposes a method for training robust convolutional networks by penalizing the predictive power of the local representations learned by earlier layers. Intuitively, our networks are forced to discard predictive signals such as color and texture that can be gleaned from local receptive fields and to rely instead on the global structures of the image. Across a battery of synthetic and benchmark domain adaptation tasks, our method confers improved generalization out of the domain. Additionally, to evaluate cross-domain transfer, we introduce ImageNet-Sketch, a new dataset consisting of sketch-like images that matches the ImageNet classification validation set in scale and dimension.

The ability to inferring latent psychological traits from human behavior is key to developing personalized human-interacting machine learning systems. Approaches to infer such traits range from surveys to manually-constructed experiments and games. However, these traditional games are limited because they are typically designed based on heuristics. In this paper, we formulate the task of designing behavior diagnostic games that elicit distinguishable behavior as a mutual information maximization problem, which can be solved by optimizing a variational lower bound. Our framework is instantiated by using prospect theory to model varying player traits, and Markov Decision Processes to parameterize the games. We validate our approach empirically, showing that our designed games can successfully distinguish among players with different traits, outperforming manually-designed ones by a large margin.

We present a new algorithm that significantly improves the efficiency of exploration for deep Q-learning agents in dialogue systems. Our agents explore via Thompson sampling, drawing Monte Carlo samples from a Bayes-by-Backprop neural network. Our algorithm learns much faster than common exploration strategies such as ε-greedy, Boltzmann, bootstrapping, and intrinsic-reward-based ones. Additionally, we show that spiking the replay buffer with experiences from just a few successful episodes can make Q-learning feasible when it might otherwise fail.

Many recent papers address reading comprehension, where examples consist of (question, passage, answer) tuples.Presumably, a model must combine information from both questions and passages to predict corresponding answers.However, despite intense interest in the topic, with hundreds of published papers vying for leaderboard dominance, basic questions about the difficulty of many popular benchmarks remain unanswered.In this paper, we establish sensible baselines for the bAbI, SQuAD, CBT, CNN, and Whodid-What datasets, finding that question-and passage-only models often perform surprisingly well.On 14 out of 20 bAbI tasks, passage-only models achieve greater than 50% accuracy, sometimes matching the full model.Interestingly, while CBT provides 20-sentence passages, only the last is needed for comparably accurate prediction.By comparison, SQuAD and CNN appear better-constructed.

Several recent papers investigate Active Learning (AL) for mitigating the datadependence of deep learning for natural language processing.However, the applicability of AL to real-world problems remains an open question.While in supervised learning, practitioners can try many different methods, evaluating each against a validation set before selecting a model, AL affords no such luxury.Over the course of one AL run, an agent annotates its dataset exhausting its labeling budget.Thus, given a new task, an active learner has no opportunity to compare models and acquisition functions.This paper provides a largescale empirical study of deep active learning, addressing multiple tasks and, for each, multiple datasets, multiple models, and a full suite of acquisition functions.We find that across all settings, Bayesian active learning by disagreement, using uncertainty estimates provided either by Dropout or Bayes-by-Backprop significantly improves over i.i.d.baselines and usually outperforms classic uncertainty sampling.

Knowledge Distillation (KD) consists of transferring “knowledge” from one machine learning model (the teacher) to another (the student). Commonly, the teacher is a high-capacity model with formidable performance, while the student is more compact. By transferring knowledge, one hopes to benefit from the student’s compactness, without sacrificing too much performance. We study KD from a new perspective: rather than compressing models, we train students parameterized identically to their teachers. Surprisingly, these Born-Again Networks (BANs), outperform their teachers significantly, both on computer vision and language modeling tasks. Our experiments with BANs based on DenseNets demonstrate state-of-the-art performance on the CIFAR-10 (3.5%) and CIFAR-100 (15.5%) datasets, by validation error. Additional experiments explore two distillation objectives: (i) Confidence-Weighted by Teacher Max (CWTM) and (ii) Dark Knowledge with Permuted Predictions (DKPP). Both methods elucidate the essential components of KD, demonstrating the effect of the teacher outputs on both predicted and non-predicted classes.

Faced with distribution shift between training and test set, we wish to detect and quantify the shift, and to correct our classifiers without test set labels. Motivated by medical diagnosis, where diseases (targets), cause symptoms (observations), we focus on label shift, where the label marginal p(y) changes but the conditional p(x| y) does not. We propose Black Box Shift Estimation (BBSE) to estimate the test distribution p(y). BBSE exploits arbitrary black box predictors to reduce dimensionality prior to shift correction. While better predictors give tighter estimates, BBSE works even when predictors are biased, inaccurate, or uncalibrated, so long as their confusion matrices are invertible. We prove BBSE’s consistency, bound its error, and introduce a statistical test that uses BBSE to detect shift. We also leverage BBSE to correct classifiers. Experiments demonstrate accurate estimates and improved prediction, even on high-dimensional datasets of natural images.

Following precedent in employment discrimination law, two notions of disparity are widely-discussed in papers on fairness and ML. Algorithms exhibit treatment disparity if they formally treat members of protected subgroups differently; algorithms exhibit impact disparity when outcomes differ across subgroups (even unintentionally). Naturally, we can achieve impact parity through purposeful treatment disparity. One line of papers aims to reconcile the two parities proposing disparate learning processes (DLPs). Here, the sensitive feature is used during training but a group-blind classifier is produced. In this paper, we show that: (i) when sensitive and (nominally) nonsensitive features are correlated, DLPs will indirectly implement treatment disparity, undermining the policy desiderata they are designed to address; (ii) when group membership is partly revealed by other features, DLPs induce within-class discrimination; and (iii) in general, DLPs provide suboptimal trade-offs between accuracy and impact parity. Experimental results on several real-world datasets highlight the practical consequences of applying DLPs.

Despite rapid advances in speech recognition, current models remain brittle to superficial perturbations to their inputs. Small amounts of noise can destroy the performance of an otherwise state-of-the-art model. To harden models against background noise, practitioners often perform data augmentation, adding artificially-noised examples to the training set, carrying over the original label. In this paper, we hypothesize that a clean example and its superficially perturbed counterparts shouldn't merely map to the same class - they should map to the same representation. We propose invariant-representation-learning (IRL): At each training iteration, for each training example, we sample a noisy counterpart. We then apply a penalty term to coerce matched representations at each layer (above some chosen layer). Our key results, demonstrated on the LibriSpeech dataset are the following: (i) IRL significantly reduces character error rates (CER) on both `clean' (3.3% vs 6.5%) and `other' (11.0% vs 18.1%) test sets; (ii) on several-of-domain noise settings (different from those seen during training) IRL's benefits are even more pronounced. Careful ablations confirm that our results are not simply due to shrinking activations at the chosen layers.

Dance Dance Revolution (DDR) is a popular rhythm-based video game. Players perform steps on a dance platform in synchronization with music as directed by on-screen step charts. While many step charts are available in standardized packs, players may grow tired of existing charts, or wish to dance to a song for which no chart exists. We introduce the task of learning to choreograph. Given a raw audio track, the goal is to produce a new step chart. This task decomposes naturally into two subtasks: deciding when to place steps and deciding which steps to select. For the step placement task, we combine recurrent and convolutional neural networks to ingest spectrograms of low-level audio features to predict steps, conditioned on chart difficulty. For step selection, we present a conditional LSTM generative model that substantially outperforms n-gram and fixed-window approaches.

Traditional approaches to recommendation focus on learning from large volumes of historical feedback to estimate simple numerical quantities (Will a user click on a product? Make a purchase? etc.). Natural language approaches that model information like product reviews have proved to be incredibly useful in improving the performance of such methods, as reviews provide valuable auxiliary information that can be used to better estimate latent user preferences and item properties. In this paper, rather than using reviews as an inputs to a recommender system, we focus on generating reviews as the model’s output. This requires us to efficiently model text (at the character level) to capture the preferences of the user, the properties of the item being consumed, and the interaction between them (i.e., the user’s preference). We show that this can model can be used to (a) generate plausible reviews and estimate nuanced reactions; (b) provide personalized rankings of existing reviews; and (c) recommend existing products more effectively.