Overview

Neural Collapse describes how late-training features can become compact within each class, arrange class means in a regular simplex-like pattern, and align classifier weights with those means. Researchers study when this geometry emerges using theoretical feature models, training-dynamics analyses, and diagnostics across optimizers, architectures, and data regimes. Related methods build prototype-based or contrastive losses around this geometry, with applications including long-tailed classification, few-shot learning, and test-time adaptation. A central open tension is whether collapse is a general property of trained networks or is altered or prevented by factors such as optimization, class imbalance, task structure, and the need to retain contextual information.

Latest papers

24 papers5 papers in the last four weeks, with none the four weeks before. 0.0% of all new papers.

Nov 5, 2026

Neural Collapse on Imbalanced Datasets and Open-Set Scenario

Neural collapse (NC) describes the elegant geometric structure that sample features and classifier vectors exhibit when a classification model's loss reaches its minimum. Class imbalance, however, generally distorts standard NC. In this paper, we theoretically reformulate NC on imbalanced datasets and in open-set settings. Our analysis is based on an optimization problem over class-wise averaged losses, rather than the global averaged losses used in prior NC studies. We prove that when class-wise averaged losses attain their minima, standard NC is achieved on balanced or imbalanced datasets, for both cross-entropy (CE) and binary cross-entropy (BCE) losses. We further extend the analysis to an open-set-inspired scenario, where some classes contain no training samples yet are still assigned classifier vectors in the losses. We find that in this scenario, when class-wise averaged losses reach their minima, standard NC still emerges on the known classes with training samples, while the classifier vectors of unknown classes without training samples overlap with each other and become orthogonal to those of the known classes. Finally, we analyze Softmax- and Sigmoid-based contrastive losses re-balanced by inverse sample frequency, proving that at their minima on balanced or imbalanced datasets, sample features of different classes likewise form standard NC.
Nov 5, 2026

Comparing Classification and Feature Learning by Analyzing Their Loss Functions in Normalized Spaces with Different Scale Factors

Classification and feature learning are two closely interrelated fundamental tasks. Feature learning in supervised settings is typically achieved by training classification models, under the common assumption that higher classification accuracy implies better feature quality. In this paper, we first point out that classification accuracy reflects the sample-wise local properties of a model on a dataset, whereas feature learning aims to learn features with good dataset-level global properties, such as intra-class compactness and inter-class separability. From the first principle of feature learning, we redesign the binary cross-entropy (BCE) loss, a counterpart of the cross-entropy (CE) loss used in classification. Then, in normalized feature space, we analyze the minima of the BCE and CE losses and show that both can lead to neural collapse, where sample features from different classes become maximally compact and maximally separated. More importantly, by investigating the convergence behaviors of the two losses as the scale factor γγ varies in the normalized space, we reveal how classification accuracy and feature quality change differently across spaces of different sizes. Under a fixed training strategy, a very large γγ achieves perfect classification accuracy but slows convergence toward the ideal feature geometry, whereas a very small γγ makes classification highly sensitive to small perturbations. A moderate γγ is therefore necessary for simultaneously achieving reliable classification and well-structured features. Extensive experiments validate these findings and demonstrate that BCE loss exhibits advantages in long-tail and open-set recognitions, as it better aligns with the global requirements of feature learning.
Oct 27, 2026

Full Collapse Is Suboptimal: Controlled Partial Neural Collapse for Long-Tailed Recognition

Recent long-tailed recognition (LTR) methods incorporate neural collapse (NC) and align features to a fixed Equiangular Tight Frame (ETF) as a uniform target for all classes. However, forcing tail classes to fully collapse over-contracts their samples and weakens generalization. To address this, we propose a controlled partial neural collapse, realized through two coordinated controls on a backbone: the Class-Adaptive Sparse Activation (CASA) classifier and the ETF-constrained adaptive margin (ETF-AM) loss. Rather than forcing a uniform full collapse, this regime allows head classes to collapse along the ETF while tail classes retain within-class diversity. This prevents the tail from being forced into full collapse, which is suboptimal under imbalance. To set the dimensionality of collapse, CASA induces class-related node specialization through feature-conditioned gating so that tail classes occupy their own feature directions. For the clearance axis, ETF-AM applies a frequency-adaptive margin, with smaller margins for head classes and larger for tail classes. On CIFAR100-LT, ImageNet-LT, Places-LT, and iNaturalist-2018, our method achieves the best overall accuracy over prior methods. In the empirical analysis, within-class spread increases monotonically from head to tail classes, consistent with tail classes retaining variability rather than fully collapsing.
Oct 17, 2026

An Impossibility Result for ETF Target Geometry, and an Equinorm Remedy for Imbalanced Analytic Continual Learning

Analytic continual learning (ACL) methods such as ACIL and GACL fit a ridge-regression classifier head on frozen features via the Sherman-Morrison-Woodbury identity, achieving exemplar-free, exact updates. A separate line of work, Progressive Neural Collapse (ProNC), shows standard continual classifiers benefit from a progressively expanding Simplex Equiangular Tight Frame (ETF) target. It is natural to ask whether fusing the two ideas, exact analytic updates and ETF targets, would yield a better closed-form continual learner. We show this is not the case as the fusion turns out to be identical to plain one-hot ACIL. Further, Simplex ETFs are not nested across class counts, so the literal "progressively expand the ETF" construction stops yielding an ETF as new tasks arrive, and it introduces catastrophic forgetting (4×4\times that of a classic ACIL head). We therefore separate an ETF's two defining properties, equiangularity and equinormality, and apply each on its own to the classifier weights rather than the targets, where the invariance above does not hold. Equinorm post-processing of the fitted weights recovers both tail-class and overall average accuracy across the long-tailed regime from 5:1 to 100:1 imbalance (e.g. +32.0 tail-accuracy points and +16.0 average-accuracy points at 100:1 on CIFAR-100/ViT), transferring in sign across two backbones and two datasets, at a mild cost of 5.4 average-accuracy points on CIFAR-100/ViT on near-balanced streams (below ≈3:1\approx 3{:}1). Equiangular geometry, by contrast, is exactly inert on the target side but actively harmful on the weight side, usually losing to plain one-hot. Hence, the generally useful part of Neural Collapse geometry in Analytic Continual Learning is equinorm post-processing of the weights, becoming more useful as class imbalance grows; and equiangularity is at best a no-op and induces catastrophic forgetting at worst.
Oct 15, 2026

A Closed Form Depth-Width Law for Neural Collapse in Deep Linear Networks

How the depth and width of neural networks shape learned representations is difficult to predict from the data and architecture. We use neural collapse (NC) as a probe of the representation geometry and express its first three properties, within-class variability collapse (NC1), simplex class means (NC2) and classifier alignment (NC3), through the learned feature kernel, extending a known identity for NC1 to NC2 and NC3. For deep linear networks with an exchangeable input kernel, the kernel equations yield a geometric law, exact within this theory, for the layerwise decrease of NC1. At large width NN the accumulated decrease becomes a function of the depth-to-width ratio L/NL/N, with its rate determined self-consistently by the data and architecture, yielding a scaling law for representation geometry.
Aug 20, 2026

Reliable Neural Collapse Approximation for Open-World Test-Time Adaptation

Test-Time Adaptation (TTA) methods aim to bridge the domain gap between the source and target domains. However, traditional TTA methods become ineffective when the label distribution shift occurs, a challenge commonly referred to as an open-world scenario. In this paper, we introduce a new method named Reliable Neural Collapse approximation (ReNC) for Open-World Test-Time Adaptation (OWTTA). Specifically, we leverage neural collapse as a structural prior for reliable target-domain adaptation. Guided by this prior, we justify that the pre-trained classifier weights can serve as the prototypes of the source domain. By measuring the similarity between samples and prototypes, we filter out the Out-Of-Distribution~(OOD) samples for reliable updates. Furthermore, we propose a neural collapse approximation mechanism to refine these prototypes, ensuring they can gradually adapt to the target domain while maintaining the neural collapse structure. Extensive experiments on several open-world benchmarks demonstrate the superiority of the proposed method. Our empirical analysis suggests that ReNC better preserves NC-related properties in the target domain, providing useful evidence for explaining reliable OWTTA and offering new insights for model design. Code is available at https://github.com/JiaqiLin-AI/ReNC.
Jul 14, 2026

Contrastive-Collapsed Loss for Flexible and Geometrically Optimal Embeddings and Faster Convergence

In this work, we introduce CoCo, a loss function aimed at learning normalized and well-structured representations. The proposed loss encourages intra-class collapse and inter-class contrast while preserving sufficient flexibility for neural networks to approximate geometrically optimal embeddings with large angular separation between classes. We provide a theoretical analysis positioning CoCo with respect to related objectives such as dot regression and cross-entropy, showing that the new proposed loss benefits from closer initialization to the optimal configuration, more informative gradients, and stronger incentives for class-wise representation collapse. Extensive experiments on diverse tabular datasets from the OpenML-CC18 benchmark show that CoCo achieves competitive performance with state-of-the-art methods, including kernel SVM, Random Forest, dot regression, and cross-entropy-based neural networks. In addition, both theoretical arguments and empirical analyses demonstrate that the proposal promotes tighter class clustering and faster convergence. These results highlight CoCo loss as an effective objective for learning discriminative representations while maintaining competitive predictive performance.
Jul 12, 2026

Beyond Coordinate Gauge: An Audited Protocol for Detecting Donor-Specific Functional Fingerprints after Neural Collapse

Independently trained neural networks have no shared neuron-index reference frame, so comparing them requires accounting for coordinate freedom. Neural Collapse sharpens this problem: networks converge toward a shared, low-dimensional geometry, raising the question of whether trajectory-specific functional variation remains distinguishable after convergence. We distinguish three claims - detectability, transplantability, and causal persistence - and address the first. Using five independently trained networks reconstructing Neural Collapse on MNIST, we apply a verified affine-correct alignment mapping donor heads into recipient coordinates. Donor-specific functional fingerprints remain distinguishable after recipient-level baseline correction: all 20 ordered donor-recipient pairs are correctly identified, with an exact permutation p=0.0083, robust to a leakage audit. These findings establish detectability under the test used here, but not transplantability or causal persistence. The study shows how alignment, ambiguity diagnostics, and leakage control combine to test cross-network variation in a controlled setting; whether this generalizes beyond it is open.
Jul 10, 2026

Neural Collapse Is Forbidden: Information Floors in Language Models

Within-class variance in language-model representations is commonly read as incomplete neural collapse. We argue it is allocated information storage, and that the allocation obeys a law. A one-line centering identity voids a family of simplex equiangular-tight-frame claims, including our own earlier ones; in dimensionless variance shares across 14 models, macro-category structure carries only 4-12% of representational variance and within-token context carries 79-91%, stable across a 100x parameter range. On the theory side, token-level weight decay penalizes a category in proportion to its type count, not its occurrence mass, reducing next-token prediction to an imbalanced K-class problem whose optimum orders category norms by type count. A converse floor, proved for binary categories, forces within-category dispersion to be at least proportional to the conditional mutual information I(token; context | category). The law holds: identity dispersion, not total variance, tracks this information across every tested model and partition, under a model-free estimate and even across models, where one model's information predicts another's dispersion; and over pretraining the category share overshoots, decays, and partially recovers, because the information it must carry never left.
Jun 25, 2026

Ordinal Neural Collapse as a Representation Prior for Visual Navigation

Learning robust navigation policies directly from visual observations remains a fundamental challenge in vision-based robotic navigation. In end-to-end imitation learning approaches, the visual encoder and action decoder are jointly optimized using a single action loss, which provides only an indirect supervisory signal to the encoder. This indirect supervision frequently results in the encoder learning ambiguous, action-agnostic representations. The problem is further complicated by substantial variations in scene structure and appearance across diverse environments, as well as the prevalence of visual distractors inherent to real-world navigation settings. Such action-agnostic features cause the navigation policy to produce inconsistent actions at ambiguous decision points, leading to navigation failure. To overcome these limitations, we propose ORION (Ordinal Neural Collapse for Visual Navigation), a method that explicitly organizes the encoder's representation space according to the ordinal structure of navigation actions. In the context of goal-directed navigation, ego-centric control categories from Far Left to Far Right exhibit a natural ordinal relationship in which neighboring classes share similar visual contexts, while semantically opposing classes differ substantially in appearance. We encourage class representations to be arranged sequentially along a single discriminative axis, while suppressing off-axis variance within each class. The pretrained encoder is then integrated into a diffusion-based navigation framework, and the full pipeline is fine-tuned end-to-end. Extensive experiments in both simulation and real-world settings show that ORION consistently outperforms end-to-end and neural collapse baselines in navigation success rate and goal progress, with notable gains in visually challenging scenarios such as complex multi-way intersections.
Jun 25, 2026

Structure Before Collapse: Transient semantic geometry in next-token prediction

Neural Collapse predicts that balanced one-hot classification pushes model representations to be equally far from each other; a symmetric configuration that depends only on the output label and ignores any semantic similarity in the inputs. This creates a puzzle: next-token prediction language models are trained predominantly (as context length increases) with one-hot labels: the same context is very unlikely to appear twice in training with different labels. However, they clearly learn latent structural features. That is, despite the one-hot training regime, a language model's contextual embeddings represent the fact that the next word in ''Mary broke the ___'' is likely to be filled by tokens in the latent classes of a) medium-sized, b) rigid, c) inanimate nouns. How does gradient descent find such categorical semantic structure when co-occurrence statistics collapse to one-hot sparsity, eliminating any shared next-tokens among different contexts? To investigate this tension we identify three synthetic controlled settings where inputs have latent semantic factors but are mapped to distinct one-hot labels. We find that semantic geometry emerges early in training, and that representations cluster by shared attributes despite receiving no explicit supervision to do so. This structure is transient: with sufficient capacity and time, the model eventually reaches the predicted symmetric state where all representations are equally separated. We study this phase transition through Gram matrix analysis and propose a preliminary modification to the commonly used unconstrained features model to capture the emergent semantic geometry.
Jun 8, 2026

Beyond Neural Collapse: Task-Intrinsic Geometry Governs Neural Representations in Modular Arithmetic

While neural collapse (NC) predicts that a KK-class-balanced classifier should organize terminal representations as a (K−1)(K-1)-dimensional simplex equiangular tight frame (ETF), modular addition consistently enters a different regime: networks compress to a two-dimensional cyclic geometry in which both classifier weights and token embeddings lie on circles. We refine the explanation of this phenomenon in three directions. First, we formalize a layerwise non-uniform training mechanism: downstream classifier weights are driven by dense cross-entropy gradients into a rank-2 equiangular configuration before upstream embeddings fully reorganize, and once this classifier plane forms, backpropagated feature gradients constrain embedding motion to the same plane while weight decay suppresses orthogonal components. Second, after this subspace locking, the induced in-plane dynamics admit an entropy-regularized transport interpretation on S1S^1; combined with modular-addition labels, this reduces embedding formation to phase alignment, whose minimizers are single-frequency characters of Z/PZ\mathbb{Z}/P\mathbb{Z} and hence equal-angle points on a circle. Third, we quantify why this solution prevails over NC: a simplex ETF gains only an O(1)O(1) advantage in cross-entropy, whereas the cyclic rank-2 solution enjoys a Θ(K)Θ(K) advantage under Schatten or weight-decay surrogates, yielding a critical threshold λcrit=Θ(1/K)λ_{\mathrm{crit}} = Θ(1/K). Our results explain both why classifier weights move first and why embeddings subsequently align with them, showing that grokking on modular arithmetic is governed not by maximal separation alone but by a task-structured trade-off between separation, symmetry, and complexity.
Jun 3, 2026

Optical-Guided Neural Collapse for SAR Few-Shot Class Incremental Learning

Few-shot class-incremental learning (FSCIL) in synthetic aperture radar imagery presents unique challenges due to severe data scarcity and SAR-specific variability. In particular, strong azimuth sensitivity in SAR induces large intra-class variation and inter-class confusion, and FSCIL sequential updates further lead to catastrophic forgetting of previously learned classes. Inspired by neural collapse, we propose an optical-guided SAR FSCIL framework, which derives orthogonal feature subspaces from a data-rich optical ATR dataset and uses them as geometric priors to guide SAR feature learning. SAR features are projected onto these orthogonal subspaces via principal angle constraints, effectively transferring discriminative structure from the optical to the SAR domain. Specifically, our projection loss and the classifier loss optimized with a frozen simplex-ETF geometry jointly induce neural collapse by concentrating features around class means while maintaining large inter-class angles. We evaluate the approach on a benchmark comprising an optical ATR dataset and a SAR ATR dataset with 24 target classes, organized into a base training session and seven incremental sessions. Compared with recent FSCIL methods including NCFSCIL and so on, our method achieves the highest final accuracy and a favorable trade-off between final performance and performance degradation. Moreover, neural collapse metrics show improved intra-class compactness and inter-class separability, indicating that the learned features more closely approximate the ideal simplex-ETF geometry.
May 29, 2026

The role of class encoding in neural collapse

Neural collapse is a structural property of the last-hidden-layer activations in neural network classification models, when trained beyond a zero classification error. In this work, we explore the role of label encoding in neural collapse by relying on the unrestricted feature model with mean squared error training loss. We demonstrate that, for one-hot encoded labels and balanced data, the uncentered mean features associated with each class transition from a simplex equiangular tight frame to an orthogonal frame when increasing the bias regularization coefficient associated with the final classifier. These structures are reminiscent of the orthogonal frame structure of one-hot encoded labels. For any arbitrary encoding, we also show that the final classifier's bias aims at centering the labels, compensating for the discrepancy between the global mean of the labels and the origin. We further discuss the role of the encoding in other neural collapse properties.
May 21, 2026

The Implicit Bias of Depth: From Neural Collapse to Softmax Codes

Neural collapse (NC) describes the structured geometry that emerges in the features and weights of trained classifiers. Recent theory suggests NC can be suboptimal in deep architectures, attributing this to an explicit low-rank bias from L2 regularization. We study the deep unconstrained feature model (UFM)-equivalent to a deep linear network with orthogonal inputs-trained without regularization, to isolate how gradient descent and depth alone shape NC. We show that depth induces an implicit low-rank bias: low-rank matrices propagate norm more efficiently through successive multiplications, promoting low-rank alternatives to NC. These alternatives, we argue, correspond to softmax codes: max-margin solutions previously found in width-bottlenecked networks. Analyzing training dynamics under spectral initialization, we identify an early-time repulsion among singular values that drives low-rank emergence, and characterize how depth shrinks NC's basin of attraction. Finally, we show that some effects act in the opposite direction: for randomly initialized networks, increasing width biases training toward higher-rank solutions. Our results provide the first asymptotic and dynamic characterization of implicit bias in deep UFMs trained with unregularized multiclass cross-entropy.
May 19, 2026

Neural Collapse by Design: Learning Class Prototypes on the Hypersphere

Supervised classification has a theoretical optimum, Neural Collapse (NC), yet neither of its two dominant paradigms reaches it in practice. Cross entropy (CE) leaves radial degrees of freedom unconstrained and converges to a degenerate geometry, while supervised contrastive learning (SCL) drives features toward NC during pretraining but discards this structure in a post hoc linear probing phase. We show that both paradigms are different appearances of the same method that contrasts prototypes on the unit hypersphere, and that closing the gap requires fixing each at its point of failure. From the CE side, we propose NTCE and NONL, two normalized losses that import contrastive optimization's missing ingredients into classifier learning: a large effective negative set and decoupled alignment and uniformity terms. From the SCL side, we prove that SCL's objective already optimizes throughout training for a principled classifier whose weights are the class mean embeddings, making linear probing both redundant and harmful. Empirically, on four benchmarks including ImageNet-1K, NTCE and NONL surpass CE accuracy, closely approximate NC (≥95%\geq 95\%), and match CE's converged NC on 4/5 metrics in under 7.5%7.5\% of its iterations, while SCL with fixed prototypes matches linear probing without the hours-long classifier training phase. The learned geometry yields +5.5%+5.5\% mean relative improvement in transfer learning, up to +8.7%+8.7\% under severe class imbalance, and improved robustness to corruptions on ImageNet-C. Our work recasts supervised learning as prototype learning on the hypersphere, with NC reached by design.
May 11, 2026

Optimal Representations for Generalized Contrastive Learning with Imbalanced Datasets

In this paper, we provide a computable characterization of the geometry of optimal representations in Contrastive Learning (CL) when the classes are imbalanced. When classes are balanced and the representation dimension is greater than the number of classes, it is well-known that the optimal representations exhibit Neural Collapse (NC), i.e., representations from the same class collapse to their class means and the class means form an Equiangular Tight Frame (ETF). For imbalanced classes and a large, generalized family of CL losses, we prove that the optimal representations of all samples from the same class collapse to their class means and their geometry exhibits an angular symmetry structure that is determined by the relative class proportions. In general, we show that the geometry can be determined by solving a convex optimization problem. Exploiting this symmetry structure, we analytically investigate a special case where class imbalance is extreme and prove that CL exhibits a phenomenon called Minority Collapse (MC) where all samples from the minority classes (classes with small probabilities) collapse into a single vector, whenever the class imbalance exceeds a threshold, which in turn depends on the regularity properties of the CL loss used and on the number of negative samples. Numerical results are provided to illustrate these phenomena and corroborate the theoretical results. We conclude by identifying a number of open problems.
May 11, 2026

Flag Varieties: A Geometric Framework for Deep Network Alignment

Alignment, the tendency of adjacent weight matrices in deep networks to develop compatible subspace orientations, underlies gradient flow, Neural Collapse, and representation similarity across architectures. Despite extensive empirical documentation, these phenomena have resisted unified theoretical treatment: existing explanations are post-hoc, each fitted to a specific observation with whatever mathematics is at hand. We reverse this direction by deriving the mathematical structure that layerwise alignment inherently demands. Using geometric invariant theory, we prove that alignment geometry has a canonical closed, polystable stratum given by a flag variety, and that subspace intersection dimension is its unique reparameterization-invariant observable, establishing that subspace metrics are not empirical conventions but mathematical necessities. This unified framework yields two dynamical consequences: ridge regularization drives subspace alignment at an exponential rate set by weight decay, whereas nonlinear activations induce a commutator obstruction to exact basis alignment, generically present in nonlinear networks and absent in linear ones. Together these give a geometric explanation of the Level-2/3 hierarchy in Neural Collapse from first principles rather than post-hoc analysis. The commutator magnitude and head subspace overlap further serve as weight-space windows into internal alignment structure, requiring no forward passes. Experiments on multilayer perceptrons, residual networks, and pretrained language models support the proposed diagnostics and delineate their scope.
May 3, 2026

How Label Imbalance Shapes Geometry: A General Spectral Analysis of Multi-Label Neural Collapse

This work investigates the phenomenon of Neural Collapse (NC) in multi-label classification, extending its conceptual framework from multi-class learning to general correlated and imbalanced multi-label settings. Although recent studies have identified a ''tag-wise averaging'' structure for multi-label features, this view relies on implicit assumptions of label balance and combinatorial symmetry. Consequently, it fails to account for the geometrical distortions caused by intrinsic label correlations and data imbalance, which are common in practice. We resolve the multiplicity-one imbalance conjecture raised by Li et al. (2024), showing that higher-multiplicity prototypes obey a class-frequency-weighted synthesis rule rather than uniform averaging. To address this, we propose a rigorous spectral-control framework to analyze the terminal phase of multi-label learning under general imbalanced conditions. We introduce the label covariance spectrum κmκ_m, a scalar controlling the distribution-dependent lower-bound geometry, derived from the second-order moment matrix of the label distribution. Contrary to the averaging perspective, our analysis reveals that the centered label covariance spectrum controls the stability of terminal geometry by quantifying the weakest centered inter-class contrast directions. We prove that the classical Tag-wise Averaging emerges only as a special case under perfect orthogonality. Numerical experiments on synthetic distributions validate our theoretical bounds. This work resolves the scaled-average aspect of the imbalance conjecture and establishes a unifying theoretical framework that extends Neural Collapse to complex, imbalanced multi-label settings.
Mar 31, 2026

Neural Collapse Dynamics: Depth, Activation, Regularisation, and Feature Norm Threshold

Neural collapse (NC) -- the convergence of penultimate-layer features to a simplex equiangular tight frame -- is well understood at equilibrium, but the dynamics governing its onset remain poorly characterised. We identify a simple and predictive regularity: NC occurs when the mean feature norm reaches a model-dataset-specific critical value, fn*, that is largely invariant to training conditions. This value concentrates tightly within each (model, dataset) pair (CV < 8%); training dynamics primarily affect the rate at which fn approaches fn*, rather than the value itself. In standard training trajectories, the crossing of fn below fn* consistently precedes NC onset, providing a practical predictor with a mean lead time of 62 epochs (MAE 24 epochs). A direct intervention experiment confirms fn* is a stable attractor of the gradient flow -- perturbations to feature scale are self-corrected during training, with convergence to the same value regardless of direction (p>0.2). Completing the (architecture)x(dataset) grid reveals the paper's strongest result: ResNet-20 on MNIST gives fn* = 5.867 -- a +458% architecture effect versus only +68% on CIFAR-10. The grid is strongly non-additive; fn* cannot be decomposed into independent architecture and dataset contributions. Four structural regularities emerge: (1) depth has a non-monotonic effect on collapse speed; (2) activation jointly determines both collapse speed and fn*; (3) weight decay defines a three-regime phase diagram -- too little slows, an optimal range is fastest, and too much prevents collapse; (4) width monotonically accelerates collapse while shifting fn* by at most 13%. These results establish feature-norm dynamics as an actionable diagnostic for predicting NC timing, suggesting that norm-threshold behaviour is a general mechanism underlying delayed representational reorganisation in deep networks.
Mar 21, 2026

Neural collapse in the orthoplex regime

When training a neural network for classification, the feature vectors of the training set are known to collapse to the vertices of a regular simplex, provided the dimension dd of the feature space and the number nn of classes satisfies n≤d+1n\leq d+1. This phenomenon is known as neural collapse. For other applications like language models, one instead takes n≫dn\gg d. Here, the neural collapse phenomenon still occurs, but with different emergent geometric figures. We characterize these geometric figures in the orthoplex regime where d+2≤n≤2dd+2\leq n\leq 2d. The techniques in our analysis primarily involve Radon's theorem and convexity.
Mar 3, 2026

Directional Neural Collapse Explains Few-Shot Transfer in Self-Supervised Learning

Frozen self-supervised representations often transfer well with only a few labels across many semantic tasks. We argue that a single geometric quantity, directional CDNV (decision-axis variance), sits at the core of two favorable behaviors: strong few-shot transfer within a task, and low interference across many tasks. We show that both emerge when variability along class-separating directions is small. First, we prove sharp non-asymptotic multiclass generalization bounds for downstream classification whose leading term is the directional CDNV. The bounds include finite-shot corrections that cleanly separate intrinsic decision-axis variability from centroid-estimation error. Second, we link decision-axis collapse to multitask geometry: for independent balanced labelings, small directional CDNV across tasks forces the corresponding decision axes to be nearly orthogonal, helping a single representation support many tasks with minimal interference. Empirically, across SSL objectives, directional CDNV collapses during pretraining even when classical CDNV remains large, and our bounds closely track few-shot error at practical shot sizes. Additionally, on synthetic multitask data, we verify that SSL learns representations whose induced decision axes are nearly orthogonal. The code and project page of the paper are available at [project page].
Feb 18, 2026

Optimizer choice matters for the emergence of Neural Collapse

Neural Collapse (NC) refers to the emergence of highly symmetric geometric structures in the representations of deep neural networks during the terminal phase of training. Despite its prevalence, the theoretical understanding of NC remains limited. Existing analyses largely ignore the role of the optimizer, thereby suggesting that NC is universal across optimization methods. In this work, we challenge this assumption and demonstrate that the choice of optimizer plays a critical role in the emergence of NC. The phenomenon is typically quantified through NC metrics, which, however, are difficult to track and analyze theoretically. To overcome this limitation, we introduce a novel diagnostic metric, NC0, whose convergence to zero is a necessary condition for NC. Using NC0, we provide theoretical evidence that NC cannot emerge under decoupled weight decay in adaptive optimizers, as implemented in AdamW. Concretely, we prove that SGD, SignGD with coupled weight decay (a special case of Adam), and SignGD with decoupled weight decay (a special case of AdamW) exhibit qualitatively different NC0 dynamics. Also, we show the accelerating effect of momentum on NC (beyond convergence of train loss) when trained with SGD, being the first result concerning momentum in the context of NC. Finally, we conduct extensive empirical experiments consisting of 3,900 training runs across various datasets, architectures, optimizers, and hyperparameters, confirming our theoretical results. This work provides the first theoretical explanation for optimizer-dependent emergence of NC and highlights the overlooked role of weight-decay coupling in shaping the implicit biases of optimizers.
Feb 10, 2026

Equilibrium contrastive learning for imbalanced image classification

Contrastive learning (CL) is a predominant technique in image classification, but they showed limited performance with an imbalanced dataset. Recently, several supervised CL methods have been proposed to promote an ideal regular simplex geometric configuration in the representation space-characterized by intra-class feature collapse and uniform inter-class mean spacing, especially for imbalanced datasets. In particular, existing prototype-based methods include class prototypes, as additional samples to consider all classes. However, the existing CL methods suffer from two limitations. First, they do not consider the alignment between the class means/prototypes and classifiers, which could lead to poor generalization. Second, existing prototype-based methods treat prototypes as only one additional sample per class, making their influence depend on the number of class instances in a batch and causing unbalanced contributions across classes. To address these limitations, we propose Equilibrium Contrastive Learning (ECL), a supervised CL framework designed to promote geometric equilibrium, where class features, means, and classifiers are harmoniously balanced under data imbalance. The proposed ECL framework uses two main components. First, ECL promotes the representation geometric equilibrium (i.e., a regular simplex geometry characterized by collapsed class samples and uniformly distributed class means), while balancing the contributions of class-average features and class prototypes. Second, ECL establishes a classifier-class center geometric equilibrium by aligning classifier weights and class prototypes. We ran experiments with three long-tailed datasets, the CIFAR-10(0)-LT, ImageNet-LT, and the two imbalanced medical datasets, the ISIC 2019 and our constructed LCCT dataset. Results show that ECL outperforms existing SOTA supervised CL methods designed for imbalanced classification.