Detects model changes in machine learning with empirical difference test.
problem Detect significant model changes between pre-change and post-change parameters.
method Constructs an empirical difference test (EDT) to approximate GLRT, with low computational complexity and false alarm constraint.
result EDT approximates GLRT and provides a method to set threshold for false alarm constraint.
Different technological domains have significantly different rates of performance improvement. Prior theory indicates that such differing rates should influence the relative speed of diffusion of the products embodying the different technologies since improvement in performance during the diffusion process increases th…
We propose a class of nonparametric two-sample tests with a cost linear in the sample size. Two tests are given, both based on an ensemble of distances between analytic functions representing each of the distributions. The first test uses smoothed empirical characteristic functions to represent the distributions, the s…
We study 'meta-dependence' in conditional independence tests across different empirical distributions.
problem Understanding the breakdown of conditional independence properties in finite data.
method Geometric intuition and information projections to measure meta-dependence between conditional independences.
result We provide a measure of meta-dependence that consolidates findings across synthetic and real-world data.
Given two networks with the same training loss on a dataset, when would they have drastically different test losses and errors? Better understanding of this question of generalization may improve practical applications of deep networks. In this paper we show that with cross-entropy loss it is surprisingly simple to ind…
Improved KSD test for better detection of differences in distributions.
problem Low power of KSD test when distributions have same modes but different mixing proportions.
method Perturb the observed sample using Markov transition kernels to improve KSD test power.
result Perturbed KSD test can lead to substantially higher power than the original KSD test.
SGD-trained models' disagreement predicts test error.
problem Estimating test error of deep networks.
method Empirical testing and theoretical analysis of SGD ensembles.
result SGD ensembles' disagreement correlates with test error.
Two semimetrics on probability distributions are proposed, given as the sum of differences of expectations of analytic functions evaluated at spatial or frequency locations (i.e, features). The features are chosen so as to maximize the distinguishability of the distributions, by optimizing a lower bound on test power f…
The paper explores how machine learning models generalize when the true distribution differs from the training distribution.
problem Understanding generalization beyond the training distribution in machine learning.
method Study through information measures, focusing on mutual information between input samples and representations.
result Bounding the testing gap with high probability using mutual information between input samples and representations.
The paper improves asymmetric causality tests by addressing inefficiencies and statistical significance issues.
problem Inefficiencies and statistical significance issues in asymmetric causality tests.
method Improved asymmetric causality tests via partial cumulative sums for positive and negative components, explicitly testing differences between causal parameters.
result Efficiently tested hypotheses on asymmetric causal interaction between financial markets.
Improved A/B testing by leveraging system similarities.
problem Traditional A/B testing ignores potential system similarities.
method Off-policy estimation to exploit system propensities.
result Improved A/B testing estimators achieve better accuracy.
New test assesses shared brain activity across different cognitive modalities.
problem Determining if different cognitive modalities use overlapping neural representations.
method Formulated a statistical hypothesis testing approach using permutation testing.
result New test (CMPT) has greater statistical power than cross-modal decoding while maintaining low Type I errors.
Calibration without labels in multiple testing
problem Interpretable error probabilities in large-scale hypothesis testing
method Constructing pseudo-labels from spacings of ordered p-values result Finding that q-value can be severely miscalibrated We derive a new discrepancy statistic for measuring differences between two probability distributions based on combining Stein's identity with the reproducing kernel Hilbert space theory. We apply our result to test how well a probabilistic model fits a set of observations, and derive a new class of powerful goodness-o…
Study evaluates off-policy reinforcement learning methods.
problem Evaluating reinforcement learning policies without direct access to the behavior policy.
method Experimental benchmarking suite with diverse design parameters.
result Guidelines for using off-policy evaluation methods in practice.
A new measure scales MMD to assess distribution closeness.
problem Testing statistical significance of distribution closeness.
method Norm-adaptive MMD (NAMMD) for distributional discrepancy.
result NAMMD-based DCT has higher test power than MMD-based DCT.
The paper improves Fisher-Pitman tests for Poisson mixtures, detecting autism-related genes.
problem Detecting differentially expressed genes between autism and control subjects.
method Nonparametric Poisson mixtures and Fisher-Pitman permutation tests.
result The tests reveal genes missed by common methods, demonstrating rate optimality.
Kernel two-sample testing is a useful statistical tool in determining whether data samples arise from different distributions without imposing any parametric assumptions on those distributions. However, raw data samples can expose sensitive information about individuals who participate in scientific studies, which make…
The paper designs tests for comparing ranked preference data and finds significant differences.
problem Comparing pairwise comparison and ranking data in various applications.
method Developed two-sample tests for pairwise comparison and ranking data, proving upper and lower bounds.
result Upper and lower bounds show tightness of the proposed tests, and significant differences in preferences were found.
Paper uses ML to improve A/B testing for complex treatment effects.
problem Detecting treatment effects in A/B experiments with complex variables.
method Combines ML models with randomization tests for better detection of treatment effects.
result ML-assisted tests improve detection of complex treatment effects.
A new one-step method for covariate shift adaptation.
problem Real-world data often violates the assumption of same distribution for training and test samples.
method Proposes a one-step optimization approach to jointly learn the model and weights.
result The proposed method achieves a generalization error bound and is empirically effective.
Paper proposes a differentially private test for joint dependence among random vectors.
problem Detecting joint dependence among sensitive data while maintaining privacy.
method Differentially private permutation methodology for dHSIC test.
result Proposed test attains minimax optimal power across privacy regimes.
The paper explains what affects the generalization gap in visual RL with and without distractors.
problem Understanding what affects the generalization gap in visual reinforcement learning.
method Theoretical analysis and empirical evidence.
result Minimizing representation distance between training and testing environments reduces the generalization gap.
The study introduces backward baselines to distinguish past prediction from future prediction in machine learning models.
problem Differentiating between past and future prediction in machine learning models.
method Theoretical, empirical, and normative arguments support a family of simple and efficient statistical tests called backward baselines.
result The study provides a meaningful backward baseline for auditing black-box prediction systems.
Empirical study of IRMv1, an invariant risk minimization framework.
problem Learning predictors invariant to spurious correlations across different training environments.
method Extending ColoredMNIST experiment to various settings.
result IRMv1 performs better as spurious correlation varies more widely.
This paper introduces novel backtests for the risk measure Expected Shortfall (ES) following the testing idea of Mincer and Zarnowitz (1969). Estimating a regression framework for the ES stand-alone is infeasible, and thus, our tests are based on a joint regression for the Value at Risk and the ES, which allows for dif…
ASK-NN detects distribution drifts in LLM-generated text.
problem Hallucinations and artificial text in LLM-generated outputs.
method Asymmetric two-sample test based on directed k-nearest-neighbor graph.
result ASK-NN is competitive with baselines on various benchmarks.
ORCA calibrates LLMs for efficient, generalizable reasoning.
problem Miscalibration of large language models leading to inefficiencies.
method Online Reasoning Calibration (ORCA) using conformal prediction and test-time training.
result ORCA provides higher efficiency and generalization across different reasoning tasks.
A new kernel test reduces noise in MMD by focusing on leading eigen-directions.
problem Noise in trailing directional components degrades power of standard kernel two-sample tests.
method Truncate MMD spectral decomposition, retaining only leading eigen-directions.
result Our method achieves superior power and robustness, especially in high-dimensional and unbalanced settings.
Study finds cryptoasset markets inefficient due to capital reallocation frictions.
problem Inefficiency in cryptoasset markets due to capital reallocation frictions.
method Examined investments with dominant and secondary risk factors, derived equilibrium restrictions, and tested empirically.
result Empirical results strongly reject necessary equilibrium restrictions, indicating market inefficiency.
Reviews techniques for model evaluation, selection, and algorithm comparison in machine learning.
problem Ensuring correct use of model evaluation, selection, and algorithm selection techniques in machine learning.
method Reviews and discusses various techniques for model evaluation, selection, and algorithm comparison, including holdout method, bootstrap, cross-validation, and statistical tests.
result Best practices and recommendations for model evaluation, selection, and algorithm comparison are provided.
New TTP framework fuses control arms while controlling Type-I error.
problem Bias in borrowing control data from previous trials.
method Kernel two-sample testing via MMD and equivalence testing.
result Higher power than standard TTP methods while maintaining error control.
Paper presents ERM with f-divergence regularization and its properties.
problem Minimizing empirical risk with f-divergence constraints. method Introduces normalization function and solves ERM-fDR via ODE. result Characterizes difference between empirical risks and provides numerical algorithm.
Study compares empirical systemic risk with balance sheet risk in interbank networks.
problem Disentangling balance sheet risk from network effects in systemic risk.
method Generalised DebtRank dynamics and maximum-entropy approach to compare observed and expected systemic risk.
result Systemic risk levels are compatible but differ significantly during turbulent times.
LCIT tests conditional independence using latent representations.
problem Detecting conditional independencies in statistical and machine learning tasks.
method Generative framework for learning latent representations of target variables X and Y, then testing for remaining dependencies.
result LCIT outperforms state-of-the-art baselines consistently under different metrics and settings.
Analyzes double descent in binary classification models with different losses.
problem Understanding the double descent phenomenon in binary classification models.
method Analytic study of gradient descent with logistic and square losses on binary linear classification models.
result The double descent phenomenon persists but with differences compared to logistic loss.
We conduct an extensive evaluation of price jump tests based on high-frequency financial data. After providing a concise review of multiple alternative tests, we document the size and power of all tests in a range of empirically relevant scenarios. Particular focus is given to the robustness of test performance to the …
A new test for conditional independence in discretized data.
problem Testing conditional independence when only discretized observations are available.
method Proposes a conditional independence test designed for discretized observations, using bridge equations to recover latent variables' information.
result Demonstrates the effectiveness of the proposed test through theoretical and empirical validation.
TCE measures calibration error with a test-based approach.
problem Measuring calibration error of probabilistic binary classifiers.
method TCE uses a novel loss function based on a statistical test.
result TCE offers clear interpretation, consistent scale, and enhanced visual representation.
The study examines how averaging data improves model performance.
problem Understanding the generalization gap in machine learning models.
method Data averaging, covariance analysis, and stochastic gradient descent (SGD) noise modeling.
result A modified generalization gap is always non-negative for a large class of model parameter distributions.
The paper develops a new framework for detecting distributional drifts conditioned on context.
problem Detecting distributional drifts in machine learning systems when context changes.
method Develops a framework using two-sample tests for conditional distributional treatment effects.
result Demonstrates effectiveness for detecting drift in subpopulations of data.
Proposes fair classification method using Wasserstein-1 distances.
problem Ensuring fairness in classification models by protecting sensitive information.
method Minimizes Wasserstein-1 distances to enforce independence between classifier outputs and sensitive information.
result Empirical performance superior to fairness baselines on benchmark datasets.
Suitability filter detects model performance degradation in real-world deployment.
problem Ensuring model reliability in safety-critical domains without access to ground truth labels.
method Uses suitability signals to evaluate classifier performance on unlabeled user data.
result The suitability filter reliably detects performance deviations due to covariate shift.
New method integrates real and synthetic data to improve machine learning models.
problem Expensive or impractical collection of high-quality data limits machine learning.
method Weighted empirical risk minimization approach for integrating surrogate data.
result Integrating surrogate data can significantly reduce test error on the original distribution.
New algorithm for safer machine learning with different testing and training distributions.
problem Challenges in modern machine learning where training and testing distributions differ.
method First-order optimization algorithm for superquantile-based learning.
result Promising numerical results show the approach's effectiveness.
A study on optimizing data augmentation weights for improved test-time predictions.
problem Improving robustness of predictions during testing with data augmentation methods.
method A weighted Test-Time Augmentation (TTA) approach based on variational Bayesian framework to optimize weights.
result Optimizing weights suppresses unwanted data augmentations and improves prediction performance.
Study uses neural networks to detect nonlinear dynamics in short time series.
problem Challenges in testing dynamical nonlinearities in short time series.
method Recurrent neural network classification framework using raw time series data.
result Classifier accuracy is higher than 50% for chaotic processes, around 50% for nonlinearly correlated noise.
New tests for identifying the number of latent factors in short panels with small time dimensions.
problem Determining the number of latent factors in short panels with small time dimensions.
method Eigenvalue tests based on variance-covariance matrices of asset returns, with assumptions on spherical errors or instrumental variables for factor betas.
result Established asymptotic distributional results and proposed a novel statistical test for weak factors.