A new algorithm detects changepoints in labeled and unlabeled data.
problem Accurate detection of abrupt changes in partially labeled data.
method Labeled Optimal Partitioning (LOPART) algorithm that fits train labels and predicts unlabeled changepoints.
result LOPART provides more accurate predictions than existing methods in both train and test sets.
Neural networks fit fewer samples than their parameters suggest in practice.
problem Understanding the practical limitations of neural network flexibility.
method Examination of neural network optimization, parameter efficiency, and loss surfaces.
result Neural networks can only fit training sets with significantly fewer samples than their parameters suggest.
Label aggregation makes learning robust to noisy labels.
problem Learning from noisy labels.
method Label aggregation and risk consistency.
result Aggregated labels lead to stronger consistency guarantees.
Proposes a method to make statistical inferences robust in spatially dependent settings with missing at random labels.
problem Statistical inference challenges with missing at random labels and spatial dependence.
method Doubly robust estimator with cross-fit nuisances and jackknife spatial HAC variance correction.
result Asymptotically valid confidence intervals with improved finite-sample calibration.
Modern neural networks are typically trained in an over-parameterized regime where the parameters of the model far exceed the size of the training data. Such neural networks in principle have the capacity to (over)fit any set of labels including pure noise. Despite this, somewhat paradoxically, neural network models tr…
Deep models can fit noisy labels, but robustness and reliability are still issues.
problem Training deep models with noisy labels leads to unreliable uncertainty quantification.
method Analysis of conditional distribution over noisy labels and evaluation of robust loss functions.
result Strictly proper and robust loss functions preserve accuracy but do not guarantee reliability.
We introduce an exploratory study on Mutation Validation (MV), a model validation method using mutated training labels for supervised learning. MV mutates training data labels, retrains the model against the mutated data, then uses the metamorphic relation that captures the consequent training performance changes to as…
While semi-supervised learning (SSL) algorithms provide an efficient way to make use of both labelled and unlabelled data, they generally struggle when the number of annotated samples is very small. In this work, we consider the problem of SSL multi-class classification with very few labelled instances. We introduce tw…
Over-parameterized deep neural networks trained by simple first-order methods are known to be able to fit any labeling of data. Such over-fitting ability hinders generalization when mislabeled training examples are present. On the other hand, simple regularization methods like early-stopping can often achieve highly no…
We consider the problem of computing the best-fitting ReLU with respect to square-loss on a training set when the examples have been drawn according to a spherical Gaussian distribution (the labels can be arbitrary). Let opt<1 be the population loss of the best-fitting ReLU. We prove: 1. Finding a ReLU wit…
New approach for feature evolution in streaming data with limited storage.
problem Rarely-provided labels in feature evolving streams.
method Incorporates manifold regularization and a buffer to adapt to different storage budgets.
result Preserves the performance of feature evolving learning across different storage budgets.
GRASP tests goodness-of-fit for binary classifiers without parametric assumptions.
problem Assessing the fit of a binary classifier to the underlying conditional law of labels given features.
method Formulates a tolerance hypothesis testing problem and proposes a novel test called GRASP.
result Proposes GRASP and Model-X GRASP tests for assessing goodness-of-fit in finite sample settings.
Self-training improves generalization by fitting to reliable pseudo-labels or gradually improving the classification plane.
problem Understanding how self-training improves generalization in high-dimensional Gaussian mixtures.
method Analyzing iterative self-training on binary Gaussian mixtures in the asymptotic limit.
result ST improves generalization by fitting to reliable pseudo-labels or gradually improving the classification plane.
Neural networks can overfit perfectly to noisy data and then grok near-optimal generalization.
problem Neural networks' ability to overfit perfectly to noisy data and then generalize near-optimally.
method Two-layer ReLU networks trained by gradient descent on XOR cluster data.
result Neural networks can achieve perfect fit to noisy training data and then grok near-optimal generalization.
MEC improves efficiency and robustness in semi-supervised inference.
problem Efficient inference with limited labeled data and robust uncertainty quantification.
method Machine-Learning-Assisted Generalized Entropy Calibration (MEC) using cross-fitted, calibration-weighted PPI.
result MEC achieves semiparametric efficiency bounds under weaker assumptions and provides near-nominal coverage.
An active learner is given a hypothesis class, a large set of unlabeled examples and the ability to interactively query labels to an oracle of a subset of these examples; the goal of the learner is to learn a hypothesis in the class that fits the data well by making as few label queries as possible. This work addresses…
State-of-the-art results on image recognition tasks are achieved using over-parameterized learning algorithms that (nearly) perfectly fit the training set and are known to fit well even random labels. This tendency to memorize the labels of the training data is not explained by existing theoretical analyses. Memorizati…
Paper tackles noisy labels by compressing feature representations.
problem Learning with noisy labels leads to overfitting and poor generalization.
method Introduces compression inductive bias using Dropout and Nested Dropout.
result Compression helps in combating label noise and improving performance.
The study initiates a theoretical analysis of dynamic benchmarking models.
problem Lack of theoretical foundation and empirical studies in dynamic benchmarks.
method Examined two realizations of dynamic benchmarking: sequential and hierarchical dependency models.
result Sequential dynamic benchmarks show initial performance improvement but can stall after three rounds due to label noise.
DP models misspecify LF dependencies, leading to significant performance errors.
problem Misspecification of LF dependencies in DP models.
method Theoretical bounds and empirical analysis of modeling errors.
result Modeling errors can be substantial, even with sensible LF structures.
Enhances graph classification models on small datasets.
problem Over-fitting and undergeneralization on small-scale benchmark datasets.
method Data augmentation via graph structure transformation and model evolution framework.
result Average improvement of 3 - 13% accuracy on graph classification tasks.
Binary classification improves with a small fraction of corrupted labels.
problem Binary classification with corrupted labels.
method Established corruption as a form of regularization and computed upper bounds on estimation error.
result Corruption is beneficial only up to a small fraction of the total sample, scaling with the square root of the sample size.
Deep neural networks (DNNs) can easily fit a random labeling of the training data with zero training error. What is the difference between DNNs trained with random labels and the ones trained with true labels? Our paper answers this question with two contributions. First, we study the memorization properties of DNNs. O…
The study establishes SQ lower bounds for learning halfspaces and ReLUs under Gaussian marginals.
problem Agnostically learning halfspaces and ReLUs under Gaussian marginals.
method Statistical Query (SQ) lower bounds analysis.
result Proves SQ lower bounds of dpoly(1/ε) for both problems. Method reweights instances and classes to improve robustness in noisy data.
problem Improving deep learning performance in the presence of label noise.
method Formulates constrained optimization problems to assign importance weights to instances and class labels.
result Significant performance gains observed in benchmark datasets with label noise.
In many applications the process of generating label information is expensive and time consuming. We present a new method that combines active and semi-supervised deep learning to achieve high generalization performance from a deep convolutional neural network with as few known labels as possible. In a setting where a …
New geometric interpretation explains over-parameterized models and adversarial perturbations.
problem Geometric understanding of over-parameterized regression and adversarial perturbations.
method Alternative geometric interpretation of regression in feature space.
result Adversarial perturbations are a natural feature of biased models due to underlying geometry.
We consider the problem of estimating the class prior in an unlabeled dataset. Under the assumption that an additional labeled dataset is available, the class prior can be estimated by fitting a mixture of class-wise data distributions to the unlabeled data distribution. However, in practice, such an additional labeled…
We investigate the effects of the unsupervised pre-training method under the perspective of information theory. If the input distribution displays multiple views of the supervision, then unsupervised pre-training allows to learn hierarchical representation which communicates these views across layers, while disentangli…
Semi-supervised learning benefits from informative missing labels, improving classifier performance.
problem Missing labels in semi-supervised learning can be informative, improving classifier performance.
method Formulates missingness as a mixture model problem and uses EM algorithm for fitting.
result Modelling informative missingness can yield a classifier with smaller expected error than a completely labelled sample.
New method uses weak labels to create valid confidence sets for predictions.
problem Lack of labeled data in machine learning models.
method Developed a conformal prediction framework to provide valid predictive confidence sets using weakly labeled data.
result New coverage definition allows for tighter and more informative (but valid) confidence sets.
SJS model predicts label shifts in multinomial datasets.
problem Predicting label shifts in multinomial datasets.
method Sparse joint shift model for dataset shift.
result Valid predictions and class prior probabilities estimates.
Several works have aimed to explain why overparameterized neural networks generalize well when trained by Stochastic Gradient Descent (SGD). The consensus explanation that has emerged credits the randomized nature of SGD for the bias of the training process towards low-complexity models and, thus, for implicit regulari…
We present a new modeling technique for solving the problem of ecological inference, in which individual-level associations are inferred from labeled data available only at the aggregate level. We model aggregate count data as arising from the Poisson binomial, the distribution of the sum of independent but not identic…
Image classification problems are typically addressed by first collecting examples with candidate labels, second cleaning the candidate labels manually, and third training a deep neural network on the clean examples. The manual labeling step is often the most expensive one as it requires workers to label millions of im…
This paper improves multi-label classification by leveraging high-order label correlations.
problem Improving accuracy in multi-label classification tasks using label correlations.
method Exploiting high-order label correlations through a supervised learning classifier system (UCS) and label powerset (LP) strategy.
result The proposed method outperforms other LP-based methods on multiple benchmark datasets.
Method converts age labels into distributions to improve speaker age estimation.
problem Label ambiguity in age labels makes precise speaker age estimation challenging.
method Converts age labels into label distributions and uses label distribution learning.
result Our method outperforms baseline methods by reducing MAE by 10% on a real-world dataset.
Study semi-supervised learning with noisy proxy covariates, deriving bounds and showing gains.
problem Learning from noisy proxy covariates with scarce labels.
method Two-stage estimator learning kernel eigenfeatures from all proxy covariates and fitting a ridge predictor on labeled data.
result Finite sample bounds show fast labeled sample rates and consistent gains over supervised and semi-supervised baselines.
Deep networks can handle noisy labels up to a certain threshold.
problem Deep learning's robustness to noisy labels.
method Applying classical statistical theory and universal consistency of DNNs.
result Certain DNNs can tolerate massive symmetric label noise up to the information-theoretic threshold.
The study analyzes how label noise affects deep learning feature learning.
problem The impact of label noise on deep learning feature learning.
method Theoretical analysis of a two-layer convolutional neural network under noisy label conditions.
result Two key stages identified: signal learning in Stage I and noise memorization in Stage II.
Neural networks learn clean data patterns first, then noisy data, leading to improved performance initially but deteriorating later.
problem Improvement in prediction error on clean data during early training of neural networks with noisy labels.
method Theoretical analysis and experiments to explore the dynamics of gradient descent and the impact of clean and noisy data.
result Neural networks prioritize learning clean data patterns first, leading to improved performance initially but deteriorating later due to diminishing gradient dominance of clean samples over noisy ones.
Noise-ignorant empirical risk minimization achieves state-of-the-art performance on noisy data.
problem Learning with noisy labels in multi-class classification problems.
method Introducing relative signal strength (RSS) to quantify transferability and applying Noise Ignorant Empirical Risk Minimization (NI-ERM).
result NI-ERM achieves state-of-the-art performance on CIFAR-N data challenge.
DoubleMatch combines pseudo-labeling with self-supervision for SSL.
problem Lack of effective use of unlabeled data in SSL.
method Combines pseudo-labeling with self-supervised loss.
result Achieves state-of-the-art accuracies on multiple datasets.
Over-parameterized models can memorize noisy labels and still generalize well, revealing a hidden structure.
problem Understanding how over-parameterized models can simultaneously memorize noisy labels and generalize well.
method Investigated through modular arithmetic tasks with label noise using two-layer neural networks.
result Over-parameterized models can achieve near-perfect test accuracy with 80% label noise by extracting an internal generalization structure.
Much recent machine learning research has been directed towards leveraging shared statistics among labels, instances and data views, commonly referred to as multi-label, multi-instance and multi-view learning. The underlying premises are that there exist correlations among input parts and among output targets, and the …
Multi-label text classification is a popular machine learning task where each document is assigned with multiple relevant labels. This task is challenging due to high dimensional features and correlated labels. Multi-label text classifiers need to be carefully regularized to prevent the severe over-fitting in the high …
Non-convex SGD learns halfspaces with adversarial label noise efficiently.
problem Agnostically learning halfspaces in adversarial label noise settings.
method Non-convex SGD optimization for halfspace learning.
result Non-convex SGD achieves misclassification error close to optimal with adversarial noise.
Deep neural networks (DNNs) have been shown to over-fit a dataset when being trained with noisy labels for a long enough time. To overcome this problem, we present a simple and effective method self-ensemble label filtering (SELF) to progressively filter out the wrong labels during training. Our method improves the tas…