Study on Neural Collapse limits in deep learning.
problem Understanding the limits of Neural Collapse in deep learning.
method Investigated Neural Collapse in the context of generalization and feature learning, refining conjectures and conducting experiments.
result Neural Collapse primarily occurs on the train set and not on the test set, suggesting it is an optimization phenomenon with unclear connections to generalization.
New findings show DNC is not optimal for deep models, revealing a low-rank bias.
problem Theoretical limitations of DNC in non-linear models and multi-class classification.
method Analysis of non-linear models of arbitrary depth in multi-class classification.
result DNC stops being optimal for DUFM when going beyond two layers or two classes, due to a low-rank bias.
Paper explains neural collapse in neural networks using a new model.
problem Understanding neural collapse in neural networks during training.
method Introducing the unconstrained layer-peeled model (ULPM) to prove gradient flow convergence to critical points of a minimum-norm separation problem.
result Proves that all critical points are strict saddle points except the global minimizers exhibiting neural collapse.
NECO detects out-of-distribution data using neural collapse properties.
problem Detecting out-of-distribution data in machine learning models.
method NECO leverages neural collapse geometric properties to identify OOD data.
result NECO achieves state-of-the-art results on OOD detection tasks.
Wide neural networks with weight decay exhibit neural collapse.
problem Proving neural collapse in wide neural networks trained with weight decay.
method Generic guarantees on neural collapse for wide networks with weight decay, proving low training error and balancedness, and bounded conditioning.
result First proof of neural collapse in end-to-end training of wide neural networks with weight decay.
Neural collapse occurs in normalized features over a Riemannian manifold.
problem Understanding neural collapse in normalized feature models.
method Simplified multi-class classification task to a nonconvex optimization problem over the Riemannian manifold, analyzing the landscape of critical points.
result The only global minimizers are neural collapse solutions, with all other critical points being strict saddles.
Study shows neural collapse is invariant to class imbalances under certain conditions.
problem Neural collapse properties are only valid for balanced data.
method Adopted UFM and introduced SELI for invariant characterization.
result Embeddings and classifiers always interpolate a simplex-encoded label matrix regardless of class imbalances.
Our research proves neural collapse in deep ResNets and transformers is globally optimal.
problem Understanding neural collapse in deep learning models.
method Analysis of deep regularized transformers and ResNets trained with cross entropy or mean squared error loss.
result Global optima of deep regularized transformers and ResNets are approximately collapsed, becoming more prominent as depth increases.
Our paper explains deep neural collapse in multiple layers.
problem Understanding deep neural collapse in multi-layered neural networks.
method Generalized unconstrained features model for deep networks.
result Deep unconstrained features model exhibits deep neural collapse.
A new loss function HUG decouples and generalizes neural collapse.
problem Neural collapse limits in deep learning models.
method Hyperspherical uniformity gap (HUG) as a unified framework.
result HUG decouples and generalizes neural collapse, improving model flexibility and robustness.
This work justifies neural collapse under MSE loss and analyzes the optimization landscape.
problem Understanding neural collapse in deep neural networks under MSE loss.
method Global landscape analysis of vanilla nonconvex MSE loss.
result The only global minimizers are neural collapse solutions.
The paper connects neural collapse and low-rank bias in networks with L2 regularization.
problem Understanding the emergence of low-rank bias and neural collapse in L2-regularized networks.
method Unified theoretical framework linking TCV and rank of weight matrices, proving global optimality of DNC1, and establishing a benign landscape property.
result Zero TCV across intermediate layers minimizes representation cost under natural architectural constraints, and DNC1 is globally optimal.
This paper examines how different loss functions affect neural network features and performance.
problem Investigating which loss function is best for deep neural networks.
method Examining last-layer features of deep networks and drawing inspiration from the Neural Collapse phenomenon.
result All relevant loss functions (CE, LS, FL, MSE) produce equivalent features and similar performance.
Use simplified layerwise linear models to understand neural dynamics.
problem Complex neural network dynamics are hard to grasp.
method Apply simplified layerwise linear models to explain neural phenomena.
result Simplified models explain neural collapse, emergence, etc.
Study shows 'Ordinal Neural Collapse' in deep OR tasks, revealing simple geometric relationships.
problem Understanding neural collapse in deep Ordinal Regression tasks.
method Combining cumulative link models and Unconstrained Feature Model to investigate neural collapse.
result Demonstrates 'Ordinal Neural Collapse' (ONC) with three key properties.
Deep nets exhibit 'Neural Collapse' during training's final phase, simplifying decision-making.
problem Understanding and optimizing deep learning training phases.
method Direct measurements on three deepnet architectures across seven datasets.
result Deep nets exhibit 'Neural Collapse' during training's final phase, simplifying decision-making.
Deep learning models show bias and variance are aligned, not in trade-off.
problem The classical bias-variance trade-off in deep learning models.
method Empirical evidence and theoretical analysis of bias and variance in deep learning models.
result Squared bias is approximately equal to variance for correctly classified sample points in deep learning models.
This work explains neural collapse in shallow neural networks and its impact on generalization.
problem Understanding neural collapse in shallow neural networks and its effect on generalization.
method Analysis of two and three-layer ReLU neural networks, focusing on data dimension, sample size, and signal-to-noise ratio.
result Neural collapse occurs in shallow ReLU networks under certain conditions related to data properties and network architecture.
This paper extends neural collapse to class-imbalanced datasets using an unconstrained ReLU feature model.
problem Understanding neural collapse in class-imbalanced datasets with cross-entropy loss.
method Generalized neural collapse to class-imbalanced settings using an unconstrained ReLU feature model.
result Class-means converge to orthogonal vectors with different lengths, and classifier weights align to these vectors.
AGOP mechanism explains deep neural collapse in neural networks.
problem Explaining the rigid structure of data representations in deep neural networks.
method Introducing AGOP and Deep RFM to demonstrate DNC.
result AGOP mechanism causes deep neural collapse in neural networks.
This paper extends neural collapse to regression problems, revealing key features and structures.
problem Understanding the structure learned by deep neural networks in regression tasks.
method Established Neural Regression Collapse (NRC) across different models, analyzing feature and weight alignments.
result Deep neural regression models exhibit a collapsed feature space, aligning with target dimensions and covariances.
This paper investigates how large language models achieve neural collapse, a phenomenon linked to generalization.
problem Neural collapse in large language models under imbalanced and token-rich conditions.
method Empirical investigation of scaling and regularization effects on CLMs' progression towards neural collapse.
result Neural collapse properties develop with scale and regularization, linked to generalization in language modeling.
Bayesian deep learning faces posterior collapse due to likelihood vs. prior competition.
problem Posterior collapse in Bayesian deep learning models.
method Identified competition between likelihood and prior regularization in a linear latent variable model.
result Posterior collapse is related to neural and dimensional collapse, suggesting a broader learning issue.
We analyze neural collapse in neural networks, showing that features collapse to vertices of a Simplex ETF.
problem Understanding and optimizing the features learned in the last layer of neural networks during training.
method Simplified unconstrained feature model, studying the global optimization landscape of cross-entropy loss with weight decay.
result The global minimizers of the loss are Simplex ETFs, and other critical points are strict saddles with negative curvature.
Deep nets trained with MSE loss exhibit Neural Collapse, collapsing features and classifiers to class means.
problem Understanding Neural Collapse in MSE-trained deep nets.
method Developed a new MSE loss decomposition and introduced the central path concept.
result Exact dynamics of Neural Collapse along the central path can be predicted.
Proposes PSCs for UQ in deep nets without retraining.
problem Estimating uncertainty in deep nets with a single pass.
method Identifies sensitive, smooth intermediate layer, fits probabilistic model.
result PSCs achieve UQ and OOD detection performance matching existing methods.
This work investigates how neural collapse improves transfer learning for large-scale models.
problem Improving transfer learning for large-scale models with limited labeled data.
method Investigates neural collapse and develops a fine-tuning method using skip-connections.
result Feature collapse on downstream data correlates with higher transfer accuracy.
This paper extends neural collapse to imbalanced data under cross-entropy loss.
problem Analyzing neural collapse in deep networks with imbalanced data.
method Using the unconstrained feature model and cross-entropy loss, the paper studies neural collapse in imbalanced datasets.
result Feature vectors within the same class collapse to a single mean vector, but angles between them depend on sample size.
New insights into CE dynamics reveal how Hadamard initialization simplifies softmax.
problem Understanding the dynamics of cross-entropy training loss in deep learning.
method Analyzing a two-layer linear neural network with standard-basis vectors as inputs.
result Gradient flow on cross-entropy converges to neural collapse geometry, proving global convergence.
Language models allocate information storage, not collapsing into uniform representations.
problem Incomplete neural collapse in language model representations.
method Analyzing variance and information sharing across 14 models, proving an information floor.
result Within-class variance is allocated information storage, not collapsed into uniform representations.
This paper explores how kernel methods can explain data effects on neural collapse.
problem Understanding how data affects neural collapse in neural networks.
method Formulating NC1 as a function of kernel, specializing to NNGP and NTK, and exploring a data-aware Gaussian Process kernel.
result The NTK does not represent more collapsed features than the NNGP for Gaussian data, highlighting the limitations of data-independent kernels.
New model explains neural collapse and limits on minority classes in imbalanced datasets.
problem Understanding and predicting performance limits of deep learning models on imbalanced datasets.
method Layer-Peeled Model, a nonconvex optimization program isolating top layers and applying constraints.
result Reveals a new phenomenon called Minority Collapse that limits deep learning models on minority classes.
Deep linear networks exhibit collapsing features and classifiers across datasets.
problem Understanding the collapse of features and classifiers in deep linear networks.
method Theoretical and empirical analysis of deep linear networks with MSE and CE losses.
result Deep linear networks exhibit NC properties, collapsing features and classifiers to orthogonal vectors.
Study explains how noisyGD with DP improves feature learning despite high dimensionality.
problem Improving feature learning in differential privacy settings with noisyGD.
method Layer-peeled model in representation learning, error bound analysis, feature normalization, PCA.
result Misclassification error is independent of dimension in NC, and PCA improves testing accuracy.
New metric measures dynamical richness without relying on accuracy.
problem Lack of a reliable metric for measuring dynamical richness.
method Developed a computationally efficient, performance-independent metric based on low-rank bias.
result Metric recovers neural collapse as a special case and captures known transitions without accuracy.
Study feature representations induced by dependence between variables.
problem Learning feature representations from dependent random variables.
method Characterized sufficient and necessary conditions for dependence-induced representations, and provided a family of loss functions.
result Features learned from the family of loss functions can be expressed as the composition of a loss-dependent function and the maximal correlation function.
This paper analyzes the landscape of supervised contrastive loss in over-parameterized networks.
problem Understanding the structure of solutions in over-parameterized networks under supervised contrastive loss.
method Analytical approach using unconstrained features model (UFM) to study the solutions of SC loss minimization.
result All local minima of SC loss are global minima in over-parameterized networks, and the minimizer is unique (up to rotation).
Graph neural networks (GNNs) have become increasingly popular for classification tasks on graph-structured data. Yet, the interplay between graph topology and feature evolution in GNNs is not well understood. In this paper, we focus on node-wise classification, illustrated with community detection on stochastic block m…
Study models deep learning training dynamics using locally elastic SDEs to reveal feature separability.
problem Understanding how deep learning models separate features from different classes during training.
method Modeling deep learning training using locally elastic SDEs with a drift term reflecting backpropagation impact.
result Local elasticity in SDEs leads to linear separability of features, resulting in vanishing training loss.
Gradient descent and SGD achieve low test error in specific network weight regimes.
problem Optimizing two-layer ReLU networks with standard initialization.
method Gradient flow and stochastic gradient descent, analyzing margins and weight norms.
result Gradient descent and SGD can achieve globally maximal margins under certain constraints.
Paper proposes ARB-Loss to improve classification precision in imbalanced datasets.
problem Improving classification precision on minor classes in imbalanced datasets.
method Introduces Attraction-Repulsion-Balanced Loss (ARB-Loss) to balance gradients across different classes.
result ARB-Loss achieves state-of-the-art performance with one-stage training.
The paper analyzes how low-rank layers in neural networks improve generalization.
problem Understanding how low-rank layers affect generalization in neural networks.
method Applying Maurer's chain rule for Gaussian complexity to analyze rank and spectral norm constraints.
result Deep networks with low-rank layers achieve better generalization than those with full-rank layers.
Contrastive learning harms minority group representations, affecting downstream tasks.
problem Representation harm in contrastive learning, especially affecting minority groups.
method Causal mediation analysis and stochastic block model explanation.
result Representation harm in contrastive learning is partly responsible for allocation harm in downstream tasks.
We study regularized deep neural networks (DNNs) and introduce a convex analytic framework to characterize the structure of the hidden layers. We show that a set of optimal hidden layer weights for a norm regularized DNN training problem can be explicitly found as the extreme points of a convex set. For the special cas…
This paper finds ReLU restores symmetry in SCL under class imbalances.
problem Symmetry break in SCL under class imbalances.
method Analytical proof and experiments with ReLU activation and batch selection.
result ReLU restores symmetry in SCL-learned representations without loss in test accuracy.
Theoretical justification for deep networks' performance with regularization techniques.
problem Understanding the performance of deep networks trained with the square loss.
method Analysis of gradient flow and theoretical justification of regularization techniques.
result Convergence to solutions with smaller Frobenius norms leads to better classification error bounds.
This work analyzes the maximum-margin bias in quasi-homogeneous neural networks.
problem Analyzing the maximum-margin bias in quasi-homogeneous neural networks.
method Geometric analysis of gradient dynamics for quasi-homogeneous models.
result Gradient flow implicitly favors a subset of parameters, leading to asymmetric norm minimization.
New research shows SVM and related methods can overfit without harm in multiclass classification.
problem Understanding benign overfitting in multiclass classification.
method Analyzing three training algorithms: ERM with cross-entropy, least-squares, and one-vs-all SVM.
result All three algorithms can lead to classifiers that interpolate training data and have equal accuracy under high overparameterization.