PACE explains ViTs by modeling patch-level concept distributions, surpassing existing methods.
problem Lack of trustworthy post-hoc explanations for Vision Transformers (ViTs)
method Variational Bayesian explanation framework (PACE)
result PACE surpasses state-of-the-art methods in meeting desiderata for ViT explanations.
VITS improves Thompson Sampling for contextual bandits with efficient approximate inference.
problem Intractable posterior sampling in traditional Thompson Sampling for contextual bandits.
method VITS uses Gaussian Variational Inference for efficient posterior approximations.
result VITS achieves sub-linear regret bound similar to traditional TS.
Study proposes a statistical test for Vision Transformer's attention mechanisms.
problem ViT's attention mechanisms may focus on irrelevant regions, leading to unreliable evidence.
method Selective inference framework to quantify statistical significance of attentions as p-values.
result Proposed method enables reliable quantification of false positive detection probability of attentions.
Vision Transformers show different internal representations compared to CNNs.
problem Understanding how Vision Transformers solve image classification tasks.
method Comparative analysis of ViT and CNN architectures on image classification benchmarks.
result ViT has more uniform representations across all layers, while CNNs have more varied representations.
This paper analyzes shallow ViTs, providing sample complexity and SGD behavior insights.
problem Theoretical understanding of shallow ViTs, especially their sample complexity and SGD behavior.
method Data model with label-relevant and label-irrelevant tokens, theoretical analysis of shallow ViT training.
result Characterization of sample complexity for zero generalization error in shallow ViTs.
Mobile V-MoEs scale down ViTs for resource-constrained vision tasks.
problem Scaling down Vision Transformers for resource-constrained applications.
method Sparse Mixture-of-Experts (MoEs) applied to entire images, with a stable training procedure.
result Mobile V-MoEs achieve better performance-efficiency trade-offs than dense ViTs.
Paper compresses ViTs models by 60% with minimal accuracy loss.
problem Memory constraints in ViTs models.
method Activation-aware low-rank tensor approximations.
result 60% reduction in model size with <1% accuracy loss.
B-cos transformers explain Vision Transformers' decisions.
problem Lack of holistic explanations for transformer outputs.
method Formulate each component as dynamic linear, allowing a single linear transform for summarization.
result Bcos-ViTs are highly interpretable and competitive on ImageNet.
Theoretical analysis of vision transformers' performance with MAE and CL objectives.
problem Understanding the distinct representations learned by vision transformers with MAE and CL objectives.
method Modeling visual data distribution and analyzing ViTs training dynamics with gradient descent.
result ViTs trained with MAE objectives learn both global and local features, while CL-trained ViTs favor global features.
CDAM improves attention maps for ViTs, making them more class-sensitive.
problem Existing attention maps in ViTs lack class sensitivity.
method Class-discriminative attention maps (CDAM) that scale attention scores by class relevance.
result CDAM provides more class-sensitive explanations than existing methods.
Paper analyzes why deeper layers of ViTs perform worse on out-of-distribution tasks.
problem Performance degradation of intermediate layers in ViTs under distribution shift.
method Extensive linear probing experiments across various benchmarks and fine-grained analysis of transformer modules.
result Probing feedforward network activations yields best performance under significant distribution shift.
Paper proposes ADC framework to reduce ViT SL training communication overhead.
problem Reducing communication overhead in ViT SL training.
method Two parallel compression strategies: class-agnostic merging and token discarding.
result Significantly reduces communication overhead without sacrificing accuracy.
Regularizes attention scores in vision transformers using bootstrapping.
problem Noisy and diffused attention maps in ViT limit interpretability.
method Statistical learning techniques, bootstrapping of attention scores.
result Improves shrinkage and sparsity of attention scores.
BoC probe assesses neural network confidence coherence, revealing architecture-specific uncertainty.
problem Poor calibration and OOD detection in neural networks.
method Bag-of-Coins (BoC) probe compares softmax confidence to pairwise dominance probabilities.
result BoC reveals clear ID/OOD separation for some architectures but not others.
ConViT combines CNN and ViT strengths, improving image classification.
problem Combining the strengths of CNNs and ViTs while avoiding their limitations.
method Introducing GPSA, a form of positional self-attention with a soft convolutional inductive bias.
result ConViT outperforms DeiT on ImageNet while offering improved sample efficiency.
Improved DNN calibration without sacrificing accuracy.
problem Poor calibration of over-parametrized DNNs in safety-critical applications.
method Decoupling feature extraction and classification layers, and applying Gaussian priors.
result Significant improvement in model calibration with minimal training cost.
BayesDLL offers a PyTorch library for Bayesian deep learning with large models.
problem Bayesian inference for large-scale deep networks.
method Variational inference, MC-dropout, stochastic-gradient MCMC, Laplace approximation.
result BayesDLL can handle Vision Transformers and pre-trained model weights as priors.
EpiMer merges models by solving Fréchet mean on a Riemannian manifold.
problem Integrating knowledge from multiple models without retraining.
method EpiMer casts model merging as solving the Fréchet mean on a Riemannian manifold, restricting computation to a low-rank subspace.
result EpiMer outperforms flat-geometry methods on image classification tasks.
Deep learning models predict epileptic seizures with high accuracy.
problem Predicting epileptic seizures for better patient care.
method Developed Temporal Multi-Channel Transformer (TMC-T) and Vision Transformer (TMC-ViT) models for EEG signals.
result TMC-ViT model outperformed CNN in seizure prediction.
LeJEPA provides a scalable, theory-driven approach to self-supervised learning.
problem Lack of practical guidance and theory in JEPAs.
method Identified optimal Gaussian distribution and introduced SIGReg objective.
result LeJEPA achieves state-of-the-art performance with minimal hyperparameters and heuristics.
New method adapts neural networks without losing prior knowledge.
problem Understanding and enabling flexible adaptation of neural networks.
method Differential geometry framework, functionally invariant paths (FIP).
result Achieves comparable state-of-the-art performance on continual learning and sparsification tasks.
LDA improves image classification accuracy with fewer features.
problem Fine-grained image classification with pretrained features.
method Supervised dimensionality reduction with LDA before linear probing.
result LDA improves accuracy over full features in 11 out of 12 configurations.
Paper introduces multi-scale methods to improve CATE estimation from EO data.
problem Challenges in balancing fine-grained and contextual information in EO-based causal inference.
method Multi-Scale Representation Concatenation, combining Vision Transformer and Causal Forests.
result Multi-scale approach captures effect heterogeneity better than single-scale models.
A geometric account explains why 'The Dress' is ambiguous, predicting observable signatures in image processing.
problem Understanding and predicting ambiguity in image processing, particularly in intrinsic image decomposition.
method Geometric analysis of intrinsic image decomposition, focusing on the discontinuous switch in prior-mode sections.
result Predicted signatures in albedo Jacobian and Fernet curvature can be observed in various models and datasets.
Prodigy estimates learning rate without tuning, improving convergence.
problem Estimating optimal learning rate in adaptive methods.
method Prodigy modifies D-Adaptation to estimate distance to solution D D D . result Prodigy improves convergence rate by a factor of O ( log ( D / d 0 ) ) O(\sqrt{\log(D/d_0)}) O ( log ( D / d 0 ) ) . Enhances deep learning by boosting generalization and convergence.
problem Improving generalization and convergence in deep learning models.
method Implicit Regularization Enhancement (IRE) framework that decouples flat and sharp directions.
result IRE consistently improves generalization performance across various deep learning tasks and models.
Framework detects brain tumors robustly from MRI images.
problem Low clinical incidence of brain tumor cases makes diagnosis challenging.
method YOLOv8n for detection, DeiT for classification, PTP metric for evaluation.
result F1-score of 0.92 achieved with reduced computational resources.
Vision transformers benefit from non-smooth components in adaptation.
problem Understanding the role of non-smoothness in vision transformer adaptation.
method Theoretical analysis and extensive experiments on large-scale vision transformers.
result High plasticity of attention modules and feedforward layers leads to better finetuning performance.
Develops a robust multiclass classification method for deep image classifiers.
problem Tackles data contamination and robustness to outliers in deep image classifiers.
method Uses Distributionally Robust Optimization (DRO) with Wasserstein metric ambiguity sets and regularized learning.
result Reduces test error rate by up to 83.5% and loss by up to 91.3% in image classification tasks.
Evaluates deep learning models in histopathology for robustness and classification strategies.
problem Lack of comprehensive evaluation of histopathology models beyond accuracy.
method Developed a new methodology to evaluate models on five histopathology datasets, including vision transformers and CNNs.
result Identified insights into cancer classification strategies and robustness against stain variations.
Few-Shot Diffusion Models generate new samples from small image sets.
problem Generating new samples from a few images.
method Conditional Denoising Diffusion Probabilistic Models (DDPM) with patch-based input set information.
result FSDM can generate samples from previously unseen classes conditioned on as few as 5 samples.
Unified model trained on images and videos using masked autoencoding.
problem Training a single model for multiple visual modalities.
method Masked autoencoding on a Vision Transformer.
result Unified model achieves comparable or better performance than single-modality models.
New framework aligns latent representations over-the-air using intelligent metasurfaces.
problem Heterogeneous transmitter-receiver models produce misaligned latent representations in semantic communication.
method Intelligent metasurfaces (SIM) emulate supervised and zero-shot semantic aligners directly in the wave domain.
result SIMs achieve up to 90% task accuracy in high SNR regimes, robust to low SNR.
A new method for image translation using disentangled style and content preservation.
problem Difficulty in maintaining original content during reverse diffusion in diffusion-based image translation.
method Disentangled style and content representation using intermediate keys from ViT model, CLIP loss, semantic divergence loss, and resampling strategy.
result Outperforms state-of-the-art models in text-guided and image-guided translation tasks.
BLAST optimizes deep model inference by learning efficient matrix structures.
problem Efficient inference of large-scale deep neural networks.
method Introduces BLAST matrix for adaptive structured weights.
result Significant compression and performance improvements for various model sizes.
CutMix training technique improves spatial locality in Vision Transformers.
problem Improving spatial locality in Vision Transformers trained from scratch.
method Comparison of Baseline and Modern training protocols on CIFAR-10, CIFAR-100, and Tiny-ImageNet.
result CutMix training component significantly reduces Mean Attention Distance (MAD) in early layers of Vision Transformers.
GPA improves LLM training speed by 8.71% for Llama-160M models.
problem Training Large Language Models (LLMs) with high memory overhead and slow convergence.
method Generalized Primal Averaging (GPA) extends Nesterov's method to eliminate memory-intensive two-loop structure.
result GPA achieves up to 10.13% speedup over AdamW in training Llama-1B model.
Novel regularization for Vision Transformers improves model generalization and sparsity.
problem Improving generalization and sparsity in Vision Transformers.
method Likelihood-guided variational Ising-based regularization.
result Improved generalization and sparsity in Vision Transformers.
VL finds flatter solutions at edge of stability, matching theory with practice.
problem Understanding implicit regularization in deep learning.
method Edge of Stability framework, controlling variational posterior shape and sample number.
result VL finds even flatter solutions than gradient descent.
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.
This work characterizes benign overfitting in Vision Transformers.
problem Understanding generalization of Vision Transformers when trained to overfit.
method Gradient descent on a data distribution model, focusing on self-attention layer and softmax.
result Established a condition to distinguish between small and large test errors based on signal-to-noise ratio.
This paper improves SSL methods using ensemble techniques with data-dependent weighted losses.
problem Improving SSL performance and robustness with large unlabeled data.
method Developed a framework for weighted cross-entropy losses in ensembling SSL methods without altering the backbone.
result Our method outperforms state-of-the-art SSL methods on ImageNet-1K, especially in few-shot learning.
RS-NN predicts belief functions for classification, improving accuracy and uncertainty estimation.
problem Improving confidence in machine learning predictions for safety-critical applications.
method Random-Set Neural Network (RS-NN) using random set mathematics.
result RS-NN outperforms state-of-the-art methods in accuracy, uncertainty estimation, and OoD detection.
DeepWeightFlow generates diverse neural network weights efficiently.
problem Generating complete neural network weights efficiently and accurately.
method Flow Matching in weight space with Git Re-Basin and TransFusion.
result DeepWeightFlow generates high-accuracy neural networks without fine-tuning.
Muon replaces matrix gradient with polar factor, optimizing flat spectrum updates
problem Optimization bias in matrix updates
method Using polar factor of gradient
result Muon update maximizes entropy among bounded updates
Aux-NAS uses auxiliary labels to improve primary task performance without extra inference cost.
problem Improving primary task performance using auxiliary labels without increasing inference cost.
method Architecture-based approach with a flexible asymmetric structure for primary and auxiliary tasks, using Neural Architecture Search (NAS) to evolve networks with only primary-to-auxiliary connections.
result Achieves improved performance on multiple tasks without increasing inference cost.
A new Hopfield model reduces inefficiency in large transformer models.
problem Inefficiency in training large transformer-based models.
method Introduces an Outlier-Efficient Hopfield Model (OutEffHop) to improve model performance.
result Achieves up to 26% reduction in model output norms across four models.
Data taggants verify dataset ownership without harming models or requiring model internals.
problem Verifying dataset ownership without harming models or requiring model internals.
method Using pairs of out-of-distribution samples and random labels as secret keys, subtly altering a dataset to detect models trained on it.
result Data taggants reliably detect models trained on the protected dataset with high confidence, without compromising validation accuracy.