The study examines instance label stability in MIL classifiers trained on global image annotations.
problem Instance labels in MIL classifiers can be unstable, leading to incorrect fine-grained annotations.
method Investigated instance stability on 5 datasets, proposing an unsupervised measure.
result A performance-stability trade-off can be made when comparing MIL classifiers.
A new method finds stable labels for target data using random walks.
problem Automating the labeling of unlabeled data from a related domain.
method Random walk on a graph with stability probabilities to find stable labels.
result The method yields stable labels for target data, improving domain adaptation.
SST framework boosts GNN performance on few-labeled graph data.
problem Performance degradation of GNNs on graphs with few labeled nodes.
method Stabilized Self-Training (SST) framework for GNNs.
result SST methods achieve superior performance, especially on graphs with few labeled nodes.
SGD handles label noise with bounds improving over SGLD.
problem Label noise in non-convex optimization.
method Stochastic gradient descent with uniform dissipativity and smoothness conditions, using Wasserstein distance and algorithmic stability.
result Generalization error bounds with a rate of n−2/3, better than SGLD's n−1/2. Propagates soft labels on hypergraphs using optimal transportation.
problem Semi-supervised learning on hypergraphs.
method Wasserstein barycenters and message-passing algorithm.
result Generalization error bounds for 2-Wasserstein distance.
MAS scores cluster size consistency from points, robust to label changes.
problem Desired uniformity in cluster sizes, stability under label perturbations.
method Mass Agreement Score (MAS) measures point-centric cluster size consistency, robust to label changes.
result MAS yields similar scores for partitions with similar bulk structure, sensitive to genuine redistribution of cluster mass.
New method stratifies multi-label data for better classification performance.
problem Maintaining label space structure in multi-label data splits.
method Iterative stratification approach considering second-order relationships.
result Improves classification performance and stability of network characteristics.
Decouples homotopy quotients of generalised configuration spaces on surfaces.
problem Homological stability of generalised configuration spaces on surfaces.
method Analyzes actions of diffeomorphism groups and uses homotopy quotients.
result Decouples theorem for homology of homotopy quotients on surfaces.
Paper improves generalization bounds for noisy stochastic algorithms.
problem Improving generalization bounds for noisy stochastic algorithms.
method Introduces Exponential Family Langevin Dynamics (EFLD) and establishes data-dependent expected stability based generalization bounds.
result Sharp generalization bounds with O(1/n) sample dependence and gradient discrepancy.
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.
New stability bounds for GD in overparameterised shallow nets without NTK assumptions.
problem Generalisation and excess risk bounds for shallow neural networks.
method Oracle inequalities and stability analysis of GD without kernelisation.
result Oracle type bounds reveal GD's generalisation is controlled by an interpolating network with shortest GD path.
Increasing variance of losses improves learning with noisy labels.
problem Learning with noisy labels and the need to penalize variance of losses.
method Designing regularizers based on the label noise transition matrix to increase variance of losses.
result Increasing variance of losses significantly improves generalization ability.
Proposes a stable classifier using inflated argmax for multiclass classification.
problem Inherent instability of taking the maximizer in multiclass classification.
method Bagging for stable continuous scores, inflated argmax for stable labels.
result Inflated argmax provides necessary protection against unstable classifiers without loss of accuracy.
Two-sample tests improve on existing methods for microtubule data.
problem Testing differences between two groups of filament data.
method Optimal lifts and manifold stability theorem applied to microtubule data.
result New tests outperform existing methods on simulated and real data.
Paper proposes MDAT to stabilize domain alignment in label-scarce settings.
problem Stable and comprehensive domain alignment in label-scarce settings.
method Max-margin Domain-Adversarial Training (MDAT) with Adversarial Reconstruction Network (ARN).
result MDAT stabilizes gradient reversing and achieves strong robustness to hyper-parameters.
New models learn stable latent clusters without side info.
problem Stability of non-linear ICA representations without side information.
method Deep generative models with latent clusterings, compared to standard VAEs and auxiliary labeled models.
result Deep generative models with latent clusterings are as stable as models with side information.
AUASE embeds dynamic networks with stability guarantees for node comparison.
problem Stability in dynamic network embeddings for comparing nodes across time.
method Attributed unfolded adjacency spectral embedding (AUASE) for stable unsupervised learning.
result AUASE provides significant improvements in link prediction and node classification.
Paper improves generalization bounds for structured output prediction problems.
problem Large label sets in structured output prediction problems.
method Developed novel high-probability bounds and generalization bounds in expectation.
result Significantly improved generalization bounds with logarithmic dependency on label set size.
More unlabeled data improves adversarial robustness.
problem Adversarial robustness of neural networks.
method Risk decomposition theorem and unlabeled data optimization.
result Adversarially robust generalization can be achieved with more unlabeled data.
The abstract discusses homological stability in topological moduli spaces.
problem Homological stability in topological moduli spaces.
method Constructing canonical resolutions and introducing coefficient systems.
result Homological stability for various moduli spaces.
We prove homological stability for sequences of "oriented configuration spaces" as the number of points in the configuration goes to infinity. These are spaces of configurations of n points in a connected manifold M of dimension at least 2 which 'admits a boundary', with labels in a path-connected space X, and with an …
New surfaces in 3D manifolds are found that cannot be smoothly deformed into each other.
problem Finding surfaces in 3D manifolds that cannot be smoothly deformed into each other.
method Constructing infinite families of homotopic surfaces in closed genus-g surfaces, showing they are not smoothly image-concordant. result Closed surfaces with common framed dual sphere can be π1-injective but not smoothly image-concordant. New GANs learn clean labels from noisy data.
problem Training GANs with noisy labels.
method Incorporating a noise transition model to learn clean label distributions.
result rGANs can learn clean label distributions from noisy labels.
We employ a certain labeled finite graph, called a chart, in a closed oriented surface for describing the monodromy of a(n achiral) Lefschetz fibration over the surface. Applying charts and their moves with respect to Wajnryb's presentation of mapping class groups, we first generalize a signature formula for Lefschetz …
The paper examines stability of ReLU networks in tangent space and activation regions.
problem Stability and sensitivity of ReLU networks to small changes.
method Tangent sensitivity measure for ReLU networks, focusing on stability induced by individual examples.
result Tangent sensitivity correlates with the distribution of activation regions and generalization gap.
Ensemble methods have been shown to be an effective tool for solving multi-label classification tasks. In the RAndom k-labELsets (RAKEL) algorithm, each member of the ensemble is associated with a small randomly-selected subset of k labels. Then, a single label classifier is trained according to each combination of ele…
Improved fine-tuning with regularization and robustness for noisy labels.
problem Fine-tuning pre-trained models on small datasets can lead to overfitting and memorization.
method PAC-Bayes generalization bound analysis, layer-wise regularization, self-label-correction, label-reweighting.
result Improves performance by 1.76% on average for image classification tasks and 0.75% for few-shot classification.
The paper explains how data augmentation improves semi-supervised learning efficiency.
problem Improving accuracy from a small fraction of labeled data.
method Data augmentation induces a similarity graph, which is graph-Laplacian-regularized for downstream learning.
result A fast transductive rate of O(1/nL) is achieved, reducing the number of labels needed. A twisted quiver bundle is a set of holomorphic vector bundles over a complex manifold, labelled by the vertices of a quiver, linked by a set of morphisms twisted by a fixed collection of holomorphic vector bundles, labelled by the arrows. When the manifold is Kaelher, quiver bundles admit natural gauge-theoretic equat…
Transformers learn algorithms for in-context learning with bounds and stability analysis.
problem Understanding and formalizing in-context learning as an algorithm learning problem.
method Formalizing ICL as a multitask learning problem, deriving generalization bounds, and analyzing stability.
result Transformers can implement near-optimal algorithms for classical regression tasks with i.i.d. and dynamic data.
Let M_g^n be the moduli space of Riemann surfaces of genus g with n labeled marked points. We prove that, for g \geq 2, the cohomology groups {H^i(M_g^n;Q)}_{n=1}^{\infty} form a sequence of Sn representations which is representation stable in the sense of Church-Farb [CF]. In particular this result applied to the triv…
The paper develops bounds for multiclass semi-supervised learning with penalization.
problem Training multiclass classifiers with limited labeled data.
method Two-step process: clustering and penalized learning.
result Data-dependent generalization error bound with convergence rates.
Classifies actions of tori on manifolds up to diffeomorphisms.
problem Classifying actions of tori on manifolds up to diffeomorphisms.
method Using triples (Q, λ, c) to classify actions, where Q is a manifold-with-corners, λ is a unimodular labelling, and c is a cohomology class.
result Classifies locally standard smooth actions of T up to equivariant diffeomorphisms.
reval package selects best clustering solutions via stability-based validation.
problem Challenges in determining best clustering solutions due to lack of validation methods.
method Stability-based relative clustering validation methods.
result Determines best clustering solutions that generalize to unseen data.
PeL separates sensory interface optimization from decision learning.
problem Optimizing sensory interfaces without task-specific information.
method Formal separation of perception and decision learning, using metrics for stability, informativeness, and geometry.
result Updates preserving invariants are orthogonal to decision gradients.
Unified framework explains why overfitting is benign in interpolating learning.
problem Understanding why overfitting is benign in highly overparameterized models.
method Spectral-transport stability framework.
result Sharp benign-overfitting criterion and explicit phase-transition rates.
The paper addresses bias in fraud detection models by improving label recovery in payment networks.
problem Systematic bias in chargeback labels in payment networks.
method Formalizes the observation pipeline as a sequential missing-data problem with three stages and a corruption layer. Constructs the Sequential Triply Robust (STR) estimator to correct for all four impairments simultaneously.
result Achieves the semiparametric efficiency bound and provably dominates naive chargeback-based training in mean squared error.
Bayesian explanations are more resilient to adversarial attacks than deterministic ones.
problem Stability of saliency-based explanations under adversarial attacks in Neural Networks.
method Empirical and theoretical analysis of Bayesian vs deterministic Neural Networks.
result Bayesian explanations are more stable under adversarial perturbations and direct attacks.
Mixup improves neural network generalization and robustness.
problem Desirable behaviors like memorization and sensitivity to adversarial examples in deep neural networks.
method Trains neural network on convex combinations of pairs of examples and their labels.
result Improves generalization of state-of-the-art neural network architectures.
New method boosts GANs by creating new classes from clustering, improving sample quality.
problem Improving GAN sample quality with class information.
method Clustering representations learned by GAN to create new classes, enhancing conditional GANs.
result Generated samples reach state-of-the-art Inception scores for CIFAR-10 and STL-10 datasets.
Deep networks learn clean structure before memorizing corrupted labels, leaving a spectral signature in gradient centered scatter.
problem Deep networks' transition from learning clean structure to memorizing corrupted labels under label noise.
method Analysis of the centered scatter of per-example last-layer gradients to identify Fisher Rank Inflation.
result Fisher Rank Inflation is a spectral signature of memorization under label noise, with effective rank expanding during memorization.
Generative model uses neural flows for next-frame video generation conditioned on labels.
problem Blurriness and instability in video generation models.
method Proposes using Glow, a neural flow model, for next-frame video generation conditioned on labels.
result Glow model produces clearer and more stable videos compared to GANs.
CARVE validates clustering results using resampling and stability analysis.
problem Inconsistent and unreliable clustering results due to algorithm, preprocessing, and k sensitivity. method CARVE uses resampling-based validation and stability analysis to evaluate multiple clustering algorithms and hyperparameters.
result CARVE consistently recovers near-optimal clusterings and finer biological structure.
CasVAE outperforms supervised methods for star-galaxy classification.
problem Challenges in machine learning for astronomy data.
method Cascade Variational Auto-Encoder (CasVAE) for unsupervised star-galaxy classification.
result CasVAE outperforms baseline models in accuracy and stability.
Generative Adversarial Networks generate diverse logos from web data.
problem Designing logos is a tedious process; machine learning can automate it.
method Used synthetic labels from clustering to stabilize GAN training on multi-modal data.
result Generated high-diversity plausible logos and demonstrated interactive exploration.
S3VDC improves DC methods for scalability, stability, and simplicity.
problem Poor scalability, instability, and lack of simplicity in DC methods.
method Four algorithmic improvements: initial γ-training, periodic β-annealing, mini-batch GMM initialization, and inverse min-max transform. S3VDC incorporates all improvements. result S3VDC outperforms state-of-the-art methods on benchmark and industrial datasets.
Feedback loops amplify dataset biases, affecting future model performance.
problem Feedback loops amplify biases in datasets, risking future model reliability.
method Formalized system where model interactions are recorded and reused, analyzed for bias amplification.
result Models that behave like samples from the training distribution are more stable and calibrated.
A number of modern learning tasks involve estimation from heterogeneous information sources. This includes classification with labeled and unlabeled data as well as other problems with analogous structure such as competitive (game theoretic) problems. The associated estimation problems can be typically reduced to solvi…