Training on mixed distributions improves test performance even when components are unrelated.
problem Improving test performance with mismatched training and test distributions.
method Analyzing mixture distributions with different training and test proportions.
result Distribution shift can be beneficial, improving test performance even when components are unrelated.
The paper analyzes a neural network two-sample test using kernel analysis.
problem Determining if two datasets come from the same distribution.
method Time-analysis on a neural tangent kernel (NTK) two-sample test, extending to realistic neural network dynamics.
result Training times needed to detect deviations are well-separated in null and alternative hypothesis scenarios.
We study efficient deep learning training algorithms that process received wireless signals, if a test Signal to Noise Ratio (SNR) estimate is available. We focus on two tasks that facilitate source identification: 1- Identifying the modulation type, 2- Identifying the wireless technology and channel in the 2.4 GHz ISM…
Testing the implementation of deep learning systems and their training routines is crucial to maintain a reliable code base. Modern software development employs processes, such as Continuous Integration, in which changes to the software are frequently integrated and tested. However, testing the training routines requir…
Neural network accuracy improves with denser training samples.
problem Improving neural network accuracy on unseen test samples.
method Bounding empirical training error smoothed across activation regions and using it to discard high-risk test samples.
result Discarding high-risk test samples based on error bounds improves prediction accuracy by up to 20%.
This paper applies combinatorial testing to machine learning for robust model performance.
problem Identifying robust machine learning models using test and training sets.
method Adapting combinatorial interaction testing for machine learning, focusing on simple features.
result Combinatorial coverage can enhance model performance and robustness.
MNIST and Fashion MNIST are extremely popular for testing in the machine learning space. Fashion MNIST improves on MNIST by introducing a harder problem, increasing the diversity of testing sets, and more accurately representing a modern computer vision task. In order to increase the data quality of FashionMNIST, this …
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.
FLOPART solves peak detection by creating accurate train and test set predictions.
problem Correctly detecting peaks in sequential data.
method Dynamic programming changepoint algorithm with zero train label errors.
result FLOPART provides highly accurate predictions on both train and test sets.
The study examines how extra compute during testing affects the performance of large language models.
problem Understanding the conditions under which test-time scaling improves model performance.
method An in-context weight prediction task for linear regression was used to train transformers. The performance was analyzed under varying levels of test-time compute.
result Training transformers on diverse, relevant, and hard tasks leads to the best performance for test-time scaling.
Language models are generally trained on data spanning a wide range of topics (e.g., news, reviews, fiction), but they might be applied to an a priori unknown target distribution (e.g., restaurant reviews). In this paper, we first show that training on text outside the test distribution can degrade test performance whe…
DD algorithm tracks test error from train error without validation data.
problem Systematic generalization gap between train and test errors in modern model training.
method Decoupled descent (DD) algorithm that cancels data reuse biases via approximate message passing.
result DD algorithm rigorously demonstrates zero-cost validation and 100% data utilization.
In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a predi…
Study on continuous sequence classification with distribution uncertainty.
problem Classifying continuous sequences with varying distribution uncertainty.
method Proposes distribution-free tests for three test designs: fixed-length, sequential, and two-phase tests.
result Error probabilities decay exponentially fast for all test designs.
Adversarial training leads to large generalization gap, decomposed into bias and variance.
problem Understanding the large generalization gap in adversarially trained models.
method Bias-Variance decomposition of test risk as a function of adversarial perturbation radius.
result Bias increases monotonically with adversarial perturbation radius and is dominant in test risk.
The paper predicts loss scaling across different datasets and compute scales.
problem Predicting loss scaling across different datasets and compute scales.
method Derive shifted power law relationships between train and test losses.
result Shifted power law relationships hold for various datasets and tasks, improving prediction accuracy.
The study finds a trade-off between model size, test loss, and training loss for linear predictors.
problem Finding the optimal balance between model size, test loss, and training loss for linear predictors.
method Established an algorithm and distribution-independent trade-off using non-asymptotic analysis.
result Models with low test loss are either classical (close to noise level training loss) or modern (large number of parameters).
Unified data representation learning improves non-parametric two-sample testing.
problem Improving non-parametric two-sample testing accuracy.
method Proposes RL-TST framework combining IRs and DRs for better test power.
result RL-TST outperforms existing methods by leveraging both IRs and DRs.
New method estimates data influence efficiently by leveraging test samples.
problem Efficiently estimating influence of training data on model predictions.
method Mirrored Influence Hypothesis, forward pass for test samples.
result Significant improvement in efficiency over existing methods.
New measure assesses neural network models' functional similarity.
problem Measuring functional similarity between similar-performing neural networks.
method Robust nonparametric hypothesis testing framework.
result Proposed measure assesses neural networks' functional similarity.
Methods for combining predictions from different models in a supervised learning setting must somehow estimate/predict the quality of a model's predictions at unknown future inputs. Many of these methods (often implicitly) make the assumption that the test inputs are identical to the training inputs, which is seldom re…
New test detects when generative models memorize training data.
problem Detecting when generative models overfit by memorizing training data.
method A non-parametric three-sample test using training set, target distribution, and model-generated samples.
result The test effectively detects data-copying in various models and datasets.
New method optimizes language model performance for test-time strategies.
problem Mismatch between training objectives and test-time deployment of large language models.
method Tail-Extrapolated estimators to approximate best-of-N performance from limited training rollouts.
result Improved performance of best-of-N deployment across various models and datasets.
Adversarial training achieves optimal test error for shallow networks.
problem Achieving optimal adversarial test error for general data distributions.
method Applying new Rademacher complexity bounds and properties of optimal adversarial predictors.
result Adversarial training can achieve optimal adversarial test error for general data distributions.
Algorithm predicts with optimal loss by abstaining from uncertain test examples.
problem Predicting with training data not matching test data.
method Transductive abstention algorithm using labeled and unlabeled test examples.
result Optimal prediction loss guarantees with additional term for abstaining cost.
Machine learning theory has mostly focused on generalization to samples from the same distribution as the training data. Whereas a better understanding of generalization beyond the training distribution where the observed distribution changes is also fundamentally important to achieve a more powerful form of generaliza…
A probabilistic framework for online test-time adaptation
problem Adapting models to new data under distributional shift
method State-space modelling architecture
result Characterizing parameter learning, time evolution, prior tuning, and prediction
In the covariate shift learning scenario, the training and test covariate distributions differ, so that a predictor's average loss over the training and test distributions also differ. In this work, we explore the potential of extreme dimension reduction, i.e. to very low dimensions, in improving the performance of imp…
In real-world classification problems, the class balance in the training dataset does not necessarily reflect that of the test dataset, which can cause significant estimation bias. If the class ratio of the test dataset is known, instance re-weighting or resampling allows systematical bias correction. However, learning…
Securely trains fair models using homomorphic encryption.
problem Protecting sensitive features while testing model fairness.
method Fully homomorphic encryption for training and testing.
result Practical application to adult income data set.
Study loop corrections in random feature models affecting training and test errors.
problem Analyzing loop corrections in random feature models to understand training and test errors.
method Statistical physics and effective field theory approach to study loop corrections.
result Derived loop corrections to training error, test error, and generalization gap.
The repeated community-wide reuse of test sets in popular benchmark problems raises doubts about the credibility of reported test-error rates. Verifying whether a learned model is overfitted to a test set is challenging as independent test sets drawn from the same data distribution are usually unavailable, while other …
Introduces TPV to analyze model robustness without labels.
problem Analyzing post-training robustness of machine learning models.
method Parameter perturbations and test prediction variance (TPV) as a unifying framework.
result TPV connects various perturbations under a single lens, providing insights into model stability.
In many applications, one works with neural network models trained by someone else. For such pretrained models, one may not have access to training data or test data. Moreover, one may not know details about the model, e.g., the specifics of the training data, the loss function, the hyperparameter values, etc. Given on…
A new witness two-sample test improves data efficiency and power.
problem Nonparametric two-sample testing.
method Optimizes kernel and defines weights and basis points using training data.
result The new test is consistent, has well-controlled type-I error, and has comparable or higher power.
Multiple classifier system (MCS) has become a successful alternative for improving classification performance. However, studies have shown inconsistent results for different MCSs, and it is often difficult to predict which MCS algorithm works the best on a particular problem. We believe that the two crucial steps of MC…
New scaling laws optimize model size, training, and inference for better performance.
problem Trade-off between model size and inference cost in modern LLMs.
method Train-to-Test (T2) scaling laws that jointly optimize model size, training tokens, and inference samples. result Optimal pretraining decisions shift into overtraining regime, leading to stronger performance.
Methodology creates holdout and test/train sets for ML studies, preserving data for future research.
problem Preserving data for future research studies that are analysis-naive.
method Modification of k-fold cross-validation, randomization, and three-way split (holdout, test, training).
result Efficiently creates holdout and test/train sets without forcing.
Transformers trained on random classification tasks generalize well and can overfit without error.
problem Understanding how transformers generalize and overfit in-context.
method Analysis of implicit regularization during gradient descent training.
result Transformers can overfit without error and still generalize well.
Laboratory test results are an important and generally high dimensional component of a patient's Electronic Health Record (EHR). We train embedding representations (via Word2Vec and GloVe) for LOINC codes of laboratory tests from the EHRs of about 80,000 patients at a cancer center. To include information about lab tes…
We propose SEARNN, a novel training algorithm for recurrent neural networks (RNNs) inspired by the "learning to search" (L2S) approach to structured prediction. RNNs have been widely successful in structured prediction applications such as machine translation or parsing, and are commonly trained using maximum likelihoo…
The paper explains how neural networks learn less salient frequency components during training.
problem Understanding the grokking phenomenon in neural networks.
method Empirical frequency analysis of training data.
result Neural networks initially learn less salient frequency components of the test data.
In this work, we propose a method to reject out-of-distribution samples which can be adapted to any network architecture and requires no additional training data. Publicly available chest x-ray data (38,353 images) is used to train a standard ResNet-50 model to detect emphysema. Feature activations of intermediate laye…
New model leads to optimal test loss in sparse linear regression.
problem Sparse linear regression with low test loss despite interpolating training data.
method Developed a new parametrization of the model that combines benefits of ℓ1 and ℓ2 norms.
result Training via gradient descent leads to an interpolator with near-optimal test loss.
Study predicts SGD test loss for structured features.
problem Understanding test loss dynamics in SGD for structured data.
method Solveable model of SGD on mean square loss for arbitrary covariance structure.
result Simpler Gaussian model accurately predicts test loss of nonlinear models.
The paper tackles data misappropriation in LLMs by embedding watermarks and testing for their presence.
problem Detecting data misappropriation in LLMs trained on copyrighted data.
method Embedding watermarks, formulating as hypothesis testing, developing statistical framework, constructing test statistics, determining optimal thresholds, controlling errors, establishing asymptotic optimality.
result The proposed statistical testing framework effectively detects data misappropriation in LLMs.
Derives ideal train/test split for ridge regression in large data limit.
problem Finding optimal train/test split for ridge regression in large data scenarios.
method Mathematical derivation of optimal train/test split, considering ridge tuning parameter and asymptotic behavior.
result The optimal train/test split for ridge regression in the large data limit depends weakly on the ridge tuning parameter alpha.
Stable health predictions need deconfounding test set features.
problem Stability of predictions in health machine learning is compromised by selection biases.
method Deconfounding the test set features improves prediction stability across different environments.
result Improved stability achieved by deconfounding test set features.