New findings suggest non-contrastive learning has many bad minima, not just collapsed ones.
problem The effectiveness of non-contrastive learning in unsupervised feature learning.
method Theoretical analysis and controlled experiments on simple data models.
result Non-contrastive losses have a preponderance of non-collapsed bad minima, and these minima are not avoided during training.
Simplified non-contrastive learning avoids representation collapse.
problem Training failure modes in self-supervised learning.
method Hyperdimensional computing and inductive bias.
result The approach avoids representation collapses.
Develops a new self-supervised learning method combining contrastive and non-contrastive approaches.
problem Leveraging unlabeled data for representation learning, especially with high variance and low batch sizes.
method Converts a contrastive method (Spectral Contrastive Loss) into a non-contrastive form (MINC loss) to reduce variance and mutual information.
result MINC loss consistently improves upon the Spectral Contrastive loss baseline in learning image representations.
New method proves non-contrastive self-supervised learning learns useful features.
problem Understanding how non-contrastive self-supervised learning (NS-SL) learns useful features.
method Proved in a linear network, NS-SL learns a desirable projection matrix and reduces sample complexity. Suggested weight decay acts as an implicit threshold.
result DirectCopy, a simpler and more efficient algorithm, outperforms DirectPred on various datasets.
SupSiam and SupBYOL improve supervised representation learning with ANCL.
problem Improving supervised representation learning with ANCL.
method Proposed supervised ANCL framework leveraging labels to avoid collapse.
result Supervised ANCL improves representation learning across various datasets and tasks.
VJE learns latent representations without contrastive learning, providing probabilistic semantics.
problem Learning latent representations without contrastive signals.
method VJE maximizes a symmetric conditional evidence lower bound (ELBO) on paired encoder embeddings, using a Student-t distribution on a polar representation.
result VJE outperforms standard non-contrastive baselines in ImageNet-1K, CIFAR-10/100, and STL-10.
Feature normalization prevents collapse in non-contrastive learning dynamics.
problem Non-contrastive learning can collapse into a single point due to lack of repulsive force.
method Extended previous theory based on L2 loss to cosine loss, considering feature normalization.
result Cosine loss induces stable equilibrium, preventing collapse even with insufficient repulsive force.
New method removes contrastive loss by adding a prediction head, revealing learning mechanisms.
problem Understanding why neural networks learn competitive representations despite trivial optima.
method Empirical and theoretical analysis of a trainable, identity-initialized prediction head.
result The trainable prediction head enables learning all features, preventing dimensional collapse.
This work defines idealized SSL representations and improves existing methods.
problem Unclear characteristics of SSL representations leading to high downstream accuracies.
method Characterized ideal properties and derived necessary and sufficient conditions.
result Improved SSL methods and derived new objectives for contrastive and non-contrastive learning.
Unified framework for SSL methods linking contrastive and non-contrastive approaches.
problem Lack of theoretical foundations and design guidelines for SSL methods.
method Spectral manifold learning framework to unify SSL methods.
result Theoretical bridge between contrastive and non-contrastive methods.
Kernel-based SSL creates useful representations without labels.
problem Creating useful representations without labels using self-supervised learning.
method Derive methods for kernel-based SSL, focusing on contrastive and non-contrastive loss functions.
result Kernel-induced representations correlate related points and de-correlate unrelated ones.
Unified framework for self-supervised learning via latent distribution matching.
problem Lack of a unifying theoretical framework for diverse SSL methods.
method Casting SSL as latent distribution matching (LDM): maximizing alignment and uniformity.
result Derives a Bayesian filtering model and proves identifiable latent representations.
Proposes a framework to extract ordered eigenfunctions from contextual kernels.
problem Lack of exact spectral decomposition in existing methods.
method Modular building blocks for compatibility with contextual kernels and scalability.
result Extracted eigenfunctions provide effective importance scores for feature selection.
Model identifies causal structure from paired observational and interventional data with unknown soft interventions.
problem Identifying causal structure from observational and interventional data with unknown soft interventions.
method Proposes a scalable causal discovery model that aggregates subset-level PDAGs and applies contrastive cross-regime orientation rules.
result The model asymptotically recovers the identifiable PDAG and can orient additional edges compared to non-contrastive subset-restricted methods.
Gaussian graphical models (GGMs) are probabilistic tools of choice for analyzing conditional dependencies between variables in complex systems. Finding changepoints in the structural evolution of a GGM is therefore essential to detecting anomalies in the underlying system modeled by the GGM. In order to detect structur…
Deep learning for TOF-MRA without matched training data.
problem High-quality reconstruction from undersampled TOF-MRA.
method Two-stage unsupervised deep learning approach using cycleGAN theory.
result Exceeds performance of state-of-the-art CS-based methods.
CurvSSL improves SSL by aligning local manifold curvature.
problem Improving self-supervised learning by capturing local manifold geometry.
method CurvSSL augments Barlow Twins with a curvature-based regularizer to align and decorrelate embeddings across augmentations.
result Curvature-regularized SSL yields competitive or improved linear evaluation performance.