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. Code for reproducing our experiments can be found here: https://github.com/pratyushmaini/ssft
Real-world deployment of machine learning models is challenging when data evolves over time. And data does evolve 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 novel time-varying importance weight estimator that can detect gradual shifts in the distribution of data. Such an importance weight estimator allows the training method to selectively sample past data -- not just similar data from the past like a standard importance weight estimator would but also data that evolved in a similar fashion in the past. Our time-varying importance weight is quite general. We demonstrate different ways of implementing it that exploit some known structure in the evolution of data. We demonstrate and evaluate this approach on a variety of problems ranging from supervised learning tasks (multiple image classification datasets) where the data undergoes a sequence of gradual shifts of our design to reinforcement learning tasks (robotic manipulation and continuous control) where data undergoes a shift organically as the policy or the task changes.
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 datasets, we observe that self-pretraining rivals standard pretraining on the BookWiki corpus (despite using around $10\times$--$500\times$ 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. Our results suggest that in many scenarios, performance gains attributable to pretraining are driven primarily by the pretraining objective itself and are not always attributable to the incorporation of massive datasets. These findings are especially relevant in light of concerns about intellectual property and offensive content in web-scale pretraining data.
Despite decades of research on authorship attribution (AA) and authorship verification (AV), inconsistent dataset splits/filtering and mismatched evaluation methods make it difficult to assess the state of the art. In this paper, we present a survey of the fields, resolve points of confusion, introduce Valla that standardizes and benchmarks AA/AV datasets and metrics, provide a large-scale empirical evaluation, and provide apples-to-apples comparisons between existing methods. We evaluate eight promising methods on fifteen datasets (including distribution-shifted challenge sets) and introduce a new large-scale dataset based on texts archived by Project Gutenberg. Surprisingly, we find that a traditional Ngram-based model performs best on 5 (of 7) AA tasks, achieving an average macro-accuracy of $76.50\%$ (compared to $66.71\%$ for a BERT-based model). However, on the two AA datasets with the greatest number of words per author, as well as on the AV datasets, BERT-based models perform best. While AV methods are easily applied to AA, they are seldom included as baselines in AA papers. We show that through the application of hard-negative mining, AV methods are competitive alternatives to AA methods. Valla and all experiment code can be found here: https://github.com/JacobTyo/Valla
With COVID-19 now pervasive, identification of high-risk individuals is crucial. Using data from a major healthcare provider in Southwestern Pennsylvania, we develop survival models predicting severe COVID-19 progression. In this endeavor, we face a tradeoff between more accurate models relying on many features and less accurate models relying on a few features aligned with clinician intuition. Complicating matters, many EHR features tend to be under-coded, degrading the accuracy of smaller models. In this study, we develop two sets of high-performance risk scores: (i) an unconstrained model built from all available features; and (ii) a pipeline that learns a small set of clinical concepts before training a risk predictor. Learned concepts boost performance over the corresponding features (C-index 0.858 vs. 0.844) and demonstrate improvements over (i) when evaluated out-of-sample (subsequent time periods). Our models outperform previous works (C-index 0.844-0.872 vs. 0.598-0.810).
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 we have access to unlabeled data from multiple domains such that the label marginals $p_d(y)$ can shift across domains but the class conditionals $p(\mathbf{x}|y)$ 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 $p(d|\mathbf{x})$ suffices to identify $p_d(y)$ and $p_d(y|\mathbf{x})$ up to permutation. Thus motivated, we introduce a practical algorithm that leverages domain-discriminative models as follows: (i) push examples through domain discriminator $p(d|\mathbf{x})$; (ii) discretize the data by clustering examples in $p(d|\mathbf{x})$ space; (iii) perform non-negative matrix factorization on the discrete data; (iv) combine the recovered $p(y|d)$ with the discriminator outputs $p(d|\mathbf{x})$ to compute $p_d(y|x) \; \forall d$. With semi-synthetic experiments, we show that our algorithm can leverage domain information to improve state of the art unsupervised classification methods. We reveal a failure mode of standard unsupervised classification methods when feature-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.
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 necessary and sufficient conditions 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.
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 H\"older 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, yielding guarantees that hold simultaneously both over all H\"older risk functionals and over all hypotheses. Thus licensed to perform empirical risk minimization, we develop practical gradient-based methods for minimizing distortion risks (widely studied subset of H\"older 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.
The mechanisms by which certain training interventions, such as increasing learning rates and applying batch normalization, improve the generalization of deep networks remains a mystery. Prior works have speculated that "flatter" solutions generalize better than "sharper" solutions to unseen data, motivating several metrics for measuring flatness (particularly $\lambda_{max}$, the largest eigenvalue of the Hessian of the loss); and algorithms, such as Sharpness-Aware Minimization (SAM) [1], that directly optimize for flatness. Other works question the link between $\lambda_{max}$ and generalization. In this paper, we present findings that call $\lambda_{max}$'s influence on generalization further into question. We show that: (1) while larger learning rates reduce $\lambda_{max}$ for all batch sizes, generalization benefits sometimes vanish at larger batch sizes; (2) by scaling batch size and learning rate simultaneously, we can change $\lambda_{max}$ without affecting generalization; (3) while SAM produces smaller $\lambda_{max}$ for all batch sizes, generalization benefits (also) vanish with larger batch sizes; (4) for dropout, excessively high dropout probabilities can degrade generalization, even as they promote smaller $\lambda_{max}$; and (5) while batch-normalization does not consistently produce smaller $\lambda_{max}$, it nevertheless confers generalization benefits. While our experiments affirm the generalization benefits of large learning rates and SAM for minibatch SGD, the GD-SGD discrepancy demonstrates limits to $\lambda_{max}$'s ability to explain generalization in neural networks.
In recent years, machine learning (ML) has come to rely more heavily on crowdworkers, both for building bigger datasets and for addressing research questions requiring human interaction or judgment. Owing to the diverse tasks performed by crowdworkers, and the myriad ways the resulting datasets are used, it can be difficult to determine when these individuals are best thought of as workers, versus as human subjects. These difficulties are compounded by conflicting policies, with some institutions and researchers treating all ML crowdwork as human subjects research, and other institutions holding that ML crowdworkers rarely constitute human subjects. Additionally, few ML papers involving crowdwork mention IRB oversight, raising the prospect that many might not be in compliance with ethical and regulatory requirements. In this paper, we focus on research in natural language processing to investigate the appropriate designation of crowdsourcing studies and the unique challenges that ML research poses for research oversight. Crucially, under the U.S. Common Rule, these judgments hinge on determinations of "aboutness", both whom (or what) the collected data is about and whom (or what) the analysis is about. We highlight two challenges posed by ML: (1) the same set of workers can serve multiple roles and provide many sorts of information; and (2) compared to the life sciences and social sciences, ML research tends to embrace a dynamic workflow, where research questions are seldom stated ex ante and data sharing opens the door for future studies to ask questions about different targets from the original study. In particular, our analysis exposes a potential loophole in the Common Rule, where researchers can elude research ethics oversight by splitting data collection and analysis into distinct studies. We offer several policy recommendations to address these concerns.