Reinforcement learning with verifiable rewards (RLVR) has boosted math and coding in large language models, yet there has been little effort to extend RLVR into messier, real-world domains like forecasting. One sticking point is that outcome-based reinforcement learning for forecasting must learn from binary, delayed, and noisy rewards, a regime where standard fine-tuning is brittle. We show that outcome-only online RL on a 14B model can match frontier-scale accuracy and surpass it in calibration and hypothetical prediction market betting by adapting two leading algorithms, Group-Relative Policy Optimisation (GRPO) and ReMax, to the forecasting setting. Our adaptations remove per-question variance scaling in GRPO, apply baseline-subtracted advantages in ReMax, hydrate training with 100k temporally consistent synthetic questions, and introduce lightweight guard-rails that penalise gibberish, non-English responses and missing rationales, enabling a single stable pass over 110k events. Scaling ReMax to 110k questions and ensembling seven predictions yields a 14B model that matches frontier baseline o1 on accuracy on our holdout set (Brier = 0.193, p = 0.23) while beating it in calibration (ECE = 0.042, p < 0.001). A simple trading rule turns this calibration edge into \$127 of hypothetical profit versus \$92 for o1 (p = 0.037). This demonstrates that refined RLVR methods can convert small-scale LLMs into potentially economically valuable forecasting tools, with implications for scaling this to larger models.
Synthetic data has become an increasingly popular way to share data without revealing sensitive information. Though Membership Inference Attacks (MIAs) are widely considered the gold standard for empirically assessing the privacy of a synthetic dataset, practitioners and researchers often rely on simpler proxy metrics such as Distance to Closest Record (DCR). These metrics estimate privacy by measuring the similarity between the training data and generated synthetic data. This similarity is also compared against that between the training data and a disjoint holdout set of real records to construct a binary privacy test. If the synthetic data is not more similar to the training data than the holdout set is, it passes the test and is considered private. In this work we show that, while computationally inexpensive, DCR and other distance-based metrics fail to identify privacy leakage. Across multiple datasets and both classical models such as Baynet and CTGAN and more recent diffusion models, we show that datasets deemed private by proxy metrics are highly vulnerable to MIAs. We similarly find both the binary privacy test and the continuous measure based on these metrics to be uninformative of actual membership inference risk. We further show that these failures are consistent across different metric hyperparameter settings and record selection methods. Finally, we argue DCR and other distance-based metrics to be flawed by design and show a example of a simple leakage they miss in practice. With this work, we hope to motivate practitioners to move away from proxy metrics to MIAs as the rigorous, comprehensive standard of evaluating privacy of synthetic data, in particular to make claims of datasets being legally anonymous.
Conformal prediction is a popular framework of uncertainty quantification that constructs prediction sets with coverage guarantees. To uphold the exchangeability assumption, many conformal prediction methods necessitate an additional holdout set for parameter tuning. Yet, the impact of violating this principle on coverage remains underexplored, making it ambiguous in practical applications. In this work, we empirically find that the tuning bias - the coverage gap introduced by leveraging the same dataset for tuning and calibration, is negligible for simple parameter tuning in many conformal prediction methods. In particular, we observe the scaling law of the tuning bias: this bias increases with parameter space complexity and decreases with calibration set size. Formally, we establish a theoretical framework to quantify the tuning bias and provide rigorous proof for the scaling law of the tuning bias by deriving its upper bound. In the end, we discuss how to reduce the tuning bias, guided by the theories we developed.
Data augmentation methods, especially SoTA interpolation-based methods such as Fair Mixup, have been widely shown to increase model fairness. However, this fairness is evaluated on metrics that do not capture model uncertainty and on datasets with only one, relatively large, minority group. As a remedy, multicalibration has been introduced to measure fairness while accommodating uncertainty and accounting for multiple minority groups. However, existing methods of improving multicalibration involve reducing initial training data to create a holdout set for post-processing, which is not ideal when minority training data is already sparse. This paper uses multicalibration to more rigorously examine data augmentation for classification fairness. We stress-test four versions of Fair Mixup on two structured data classification problems with up to 81 marginalized groups, evaluating multicalibration violations and balanced accuracy. We find that on nearly every experiment, Fair Mixup \textit{worsens} baseline performance and fairness, but the simple vanilla Mixup \textit{outperforms} both Fair Mixup and the baseline, especially when calibrating on small groups. \textit{Combining} vanilla Mixup with multicalibration post-processing, which enforces multicalibration through post-processing on a holdout set, further increases fairness.
Machine learning algorithms often struggle to eliminate inherent data biases, particularly those arising from unreliable labels, which poses a significant challenge in ensuring fairness. Existing fairness techniques that address label bias typically involve modifying models and intervening in the training process, but these lack flexibility for large-scale datasets. To address this limitation, we introduce a data selection method designed to efficiently and flexibly mitigate label bias, tailored to more practical needs. Our approach utilizes a zero-shot predictor as a proxy model that simulates training on a clean holdout set. This strategy, supported by peer predictions, ensures the fairness of the proxy model and eliminates the need for an additional holdout set, which is a common requirement in previous methods. Without altering the classifier's architecture, our modality-agnostic method effectively selects appropriate training data and has proven efficient and effective in handling label bias and improving fairness across diverse datasets in experimental evaluations.
Transformer based large-language models (LLMs) display extreme proficiency with language yet a precise understanding of how they work remains elusive. One way of demystifying transformer predictions would be to describe how they depend on their context in terms of simple template functions. This paper takes a first step in this direction by considering families of functions (i.e. rules) formed out of simple N-gram based statistics of the training data. By studying how well these rulesets approximate transformer predictions, we obtain a variety of novel discoveries: a simple method to detect overfitting during training without using a holdout set, a quantitative measure of how transformers progress from learning simple to more complex statistical rules over the course of training, a model-variance criterion governing when transformer predictions tend to be described by N-gram rules, and insights into how well transformers can be approximated by N-gram rulesets in the limit where these rulesets become increasingly complex. In this latter direction, we find that for 79% and 68% of LLM next-token distributions on TinyStories and Wikipedia, respectively, their top-1 predictions agree with those provided by our N-gram rulesets.
The training data for many Large Language Models (LLMs) is contaminated with test data. This means that public benchmarks used to assess LLMs are compromised, suggesting a performance gap between benchmark scores and actual capabilities. Ideally, a private holdout set could be used to accurately verify scores. Unfortunately, such datasets do not exist for most benchmarks, and post-hoc construction of sufficiently similar datasets is non-trivial. To address these issues, we introduce a systematic methodology for (i) retrospectively constructing a holdout dataset for a target dataset, (ii) demonstrating the statistical indistinguishability of this retro-holdout dataset, and (iii) comparing LLMs on the two datasets to quantify the performance gap due to the dataset's public availability. Applying these methods to TruthfulQA, we construct and release Retro-Misconceptions, on which we evaluate twenty LLMs and find that some have inflated scores by as much as 16 percentage points. Our results demonstrate that public benchmark scores do not always accurately assess model properties, and underscore the importance of improved data practices in the field.
STAND is a data-efficient and computationally efficient machine learning approach that produces better classification accuracy than popular approaches like XGBoost on small-data tabular classification problems like learning rule preconditions from interactive training. STAND accounts for a complete set of good candidate generalizations instead of selecting a single generalization by breaking ties randomly. STAND can use any greedy concept construction strategy, like decision tree learning or sequential covering, and build a structure that approximates a version space over disjunctive normal logical statements. Unlike candidate elimination approaches to version-space learning, STAND does not suffer from issues of version-space collapse from noisy data nor is it restricted to learning strictly conjunctive concepts. More importantly, STAND can produce a measure called instance certainty that can predict increases in holdout set performance and has high utility as an active-learning heuristic. Instance certainty enables STAND to be self-aware of its own learning: it knows when it learns and what example will help it learn the most. We illustrate that instance certainty has desirable properties that can help users select next training problems, and estimate when training is complete in applications where users interactively teach an AI a complex program.
In this work we considerably improve the state-of-the-art SMT solving on first-order quantified problems by efficient machine learning guidance of quantifier selection. Quantifiers represent a significant challenge for SMT and are technically a source of undecidability. In our approach, we train an efficient machine learning model that informs the solver which quantifiers should be instantiated and which not. Each quantifier may be instantiated multiple times and the set of the active quantifiers changes as the solving progresses. Therefore, we invoke the ML predictor many times, during the whole run of the solver. To make this efficient, we use fast ML models based on gradient boosting decision trees. We integrate our approach into the state-of-the-art cvc5 SMT solver and show a considerable increase of the system's holdout-set performance after training it on a large set of first-order problems collected from the Mizar Mathematical Library.
Recently, transductive learning methods, which leverage holdout sets during training, have gained popularity for their potential to improve speed, accuracy, and fairness in machine learning models. Despite this, the composition of the holdout set itself, particularly the balance of sensitive sub-groups, has been largely overlooked. Our experiments on CIFAR and CelebA datasets show that compositional changes in the holdout set can substantially influence fairness metrics. Imbalanced holdout sets exacerbate existing disparities, while balanced holdouts can mitigate issues introduced by imbalanced training data. These findings underline the necessity of constructing holdout sets that are both diverse and representative.
Objectives: To present a publicly available torso segmentation network for large epidemiology datasets on volumetric interpolated breath-hold examination (VIBE) images. Materials & Methods: We extracted preliminary segmentations from TotalSegmentator, spine, and body composition networks for VIBE images, then improved them iteratively and retrained a nnUNet network. Using subsets of NAKO (85 subjects) and UK Biobank (16 subjects), we evaluated with Dice-score on a holdout set (12 subjects) and existing organ segmentation approach (1000 subjects), generating 71 semantic segmentation types for VIBE images. We provide an additional network for the vertebra segments 22 individual vertebra types. Results: We achieved an average Dice score of 0.89 +- 0.07 overall 71 segmentation labels. We scored > 0.90 Dice-score on the abdominal organs except for the pancreas with a Dice of 0.70. Conclusion: Our work offers a detailed and refined publicly available full torso segmentation on VIBE images.
Crowdsourced machine learning on competition platforms such as Kaggle is a popular and often effective method for generating accurate models. Typically, teams vie for the most accurate model, as measured by overall error on a holdout set, and it is common towards the end of such competitions for teams at the top of the leaderboard to ensemble or average their models outside the platform mechanism to get the final, best global model. In arXiv:2201.10408, the authors developed an alternative crowdsourcing framework in the context of fair machine learning, in order to integrate community feedback into models when subgroup unfairness is present and identifiable. There, unlike in classical crowdsourced ML, participants deliberately specialize their efforts by working on subproblems, such as demographic subgroups in the service of fairness. Here, we take a broader perspective on this work: we note that within this framework, participants may both specialize in the service of fairness and simply to cater to their particular expertise (e.g., focusing on identifying bird species in an image classification task). Unlike traditional crowdsourcing, this allows for the diversification of participants' efforts and may provide a participation mechanism to a larger range of individuals (e.g. a machine learning novice who has insight into a specific fairness concern). We present the first medium-scale experimental evaluation of this framework, with 46 participating teams attempting to generate models to predict income from American Community Survey data. We provide an empirical analysis of teams' approaches, and discuss the novel system architecture we developed. From here, we give concrete guidance for how best to deploy such a framework.
Liver tumor segmentation and classification are important tasks in computer aided diagnosis. We aim to address three problems: liver tumor screening and preliminary diagnosis in non-contrast computed tomography (CT), and differential diagnosis in dynamic contrast-enhanced CT. A novel framework named Pixel-Lesion-pAtient Network (PLAN) is proposed. It uses a mask transformer to jointly segment and classify each lesion with improved anchor queries and a foreground-enhanced sampling loss. It also has an image-wise classifier to effectively aggregate global information and predict patient-level diagnosis. A large-scale multi-phase dataset is collected containing 939 tumor patients and 810 normal subjects. 4010 tumor instances of eight types are extensively annotated. On the non-contrast tumor screening task, PLAN achieves 95% and 96% in patient-level sensitivity and specificity. On contrast-enhanced CT, our lesion-level detection precision, recall, and classification accuracy are 92%, 89%, and 86%, outperforming widely used CNN and transformers for lesion segmentation. We also conduct a reader study on a holdout set of 250 cases. PLAN is on par with a senior human radiologist, showing the clinical significance of our results.
High complexity models are notorious in machine learning for overfitting, a phenomenon in which models well represent data but fail to generalize an underlying data generating process. A typical procedure for circumventing overfitting computes empirical risk on a holdout set and halts once (or flags that/when) it begins to increase. Such practice often helps in outputting a well-generalizing model, but justification for why it works is primarily heuristic. We discuss the overfitting problem and explain why standard asymptotic and concentration results do not hold for evaluation with training data. We then proceed to introduce and argue for a hypothesis test by means of which both model performance may be evaluated using training data, and overfitting quantitatively defined and detected. We rely on said concentration bounds which guarantee that empirical means should, with high probability, approximate their true mean to conclude that they should approximate each other. We stipulate conditions under which this test is valid, describe how the test may be used for identifying overfitting, articulate a further nuance according to which distributional shift may be flagged, and highlight an alternative notion of learning which usefully captures generalization in the absence of uniform PAC guarantees.
Unsustainable fishing practices worldwide pose a major threat to marine resources and ecosystems. Identifying vessels that do not show up in conventional monitoring systems -- known as ``dark vessels'' -- is key to managing and securing the health of marine environments. With the rise of satellite-based synthetic aperture radar (SAR) imaging and modern machine learning (ML), it is now possible to automate detection of dark vessels day or night, under all-weather conditions. SAR images, however, require a domain-specific treatment and are not widely accessible to the ML community. Maritime objects (vessels and offshore infrastructure) are relatively small and sparse, challenging traditional computer vision approaches. We present the largest labeled dataset for training ML models to detect and characterize vessels and ocean structures in SAR imagery. xView3-SAR consists of nearly 1,000 analysis-ready SAR images from the Sentinel-1 mission that are, on average, 29,400-by-24,400 pixels each. The images are annotated using a combination of automated and manual analysis. Co-located bathymetry and wind state rasters accompany every SAR image. We also provide an overview of the xView3 Computer Vision Challenge, an international competition using xView3-SAR for ship detection and characterization at large scale. We release the data (\href{https://iuu.xview.us/}{https://iuu.xview.us/}) and code (\href{https://github.com/DIUx-xView}{https://github.com/DIUx-xView}) to support ongoing development and evaluation of ML approaches for this important application.
Predictive risk scores for adverse outcomes are increasingly crucial in guiding health interventions. Such scores may need to be periodically updated due to change in the distributions they model. However, directly updating risk scores used to guide intervention can lead to biased risk estimates. To address this, we propose updating using a `holdout set' - a subset of the population that does not receive interventions guided by the risk score. Balancing the holdout set size is essential to ensure good performance of the updated risk score whilst minimising the number of held out samples. We prove that this approach reduces adverse outcome frequency to an asymptotically optimal level and argue that often there is no competitive alternative. We describe conditions under which an optimal holdout size (OHS) can be readily identified, and introduce parametric and semi-parametric algorithms for OHS estimation. We apply our methods to the ASPRE risk score for pre-eclampsia to recommend a plan for updating it in the presence of change in the underlying data distribution. We show that, in order to minimise the number of pre-eclampsia cases over time, this is best achieved using a holdout set of around 10,000 individuals.
The training of neural networks is usually monitored with a validation (holdout) set to estimate the generalization of the model. This is done instead of measuring intrinsic properties of the model to determine whether it is learning appropriately. In this work, we suggest studying the training of neural networks with Algebraic Topology, specifically Persistent Homology (PH). Using simplicial complex representations of neural networks, we study the PH diagram distance evolution on the neural network learning process with different architectures and several datasets. Results show that the PH diagram distance between consecutive neural network states correlates with the validation accuracy, implying that the generalization error of a neural network could be intrinsically estimated without any holdout set.
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. Furthermore, (ii) shrinks the training set and its guarantee erodes with each re-use of the holdout set. In this paper, we introduce a method that leverages unlabeled data to produce generalization bounds. After augmenting our (labeled) training set with randomly labeled fresh examples, we train in the standard fashion. Whenever classifiers achieve low error on clean data and high error on noisy data, our bound provides a tight upper bound on the true risk. 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 provides practitioners with an option for certifying the generalization of deep nets even when unseen labeled data is unavailable and provides theoretical insights into the relationship between random label noise and generalization.
Organ morphology is a key indicator for prostate disease diagnosis and prognosis. For instance, In longitudinal study of prostate cancer patients under active surveillance, the volume, boundary smoothness and their changes are closely monitored on time-series MR image data. In this paper, we describe a new framework for forecasting prostate morphological changes, as the ability to detect such changes earlier than what is currently possible may enable timely treatment or avoiding unnecessary confirmatory biopsies. In this work, an efficient feature-based MR image registration is first developed to align delineated prostate gland capsules to quantify the morphological changes using the inferred dense displacement fields (DDFs). We then propose to use kernel density estimation (KDE) of the probability density of the DDF-represented \textit{future morphology changes}, between current and future time points, before the future data become available. The KDE utilises a novel distance function that takes into account morphology, stage-of-progression and duration-of-change, which are considered factors in such subject-specific forecasting. We validate the proposed approach on image masks unseen to registration network training, without using any data acquired at the future target time points. The experiment results are presented on a longitudinal data set with 331 images from 73 patients, yielding an average Dice score of 0.865 on a holdout set, between the ground-truth and the image masks warped by the KDE-predicted-DDFs.
While improving prediction accuracy has been the focus of machine learning in recent years, this alone does not suffice for reliable decision-making. Deploying learning systems in consequential settings also requires calibrating and communicating the uncertainty of predictions. To convey instance-wise uncertainty for prediction tasks, we show how to generate set-valued predictions from a black-box predictor that control the expected loss on future test points at a user-specified level. Our approach provides explicit finite-sample guarantees for any dataset by using a holdout set to calibrate the size of the prediction sets. This framework enables simple, distribution-free, rigorous error control for many tasks, and we demonstrate it in five large-scale machine learning problems: (1) classification problems where some mistakes are more costly than others; (2) multi-label classification, where each observation has multiple associated labels; (3) classification problems where the labels have a hierarchical structure; (4) image segmentation, where we wish to predict a set of pixels containing an object of interest; and (5) protein structure prediction. Lastly, we discuss extensions to uncertainty quantification for ranking, metric learning and distributionally robust learning.