Facebook's ResNeXt WSL models show exceptional robustness against image corruptions and adversarial attacks.
problem Image recognition model robustness against corruptions and adversarial attacks.
method Training with 1B images from Instagram and fine-tuning on ImageNet.
result ResNeXt WSL models achieve state-of-the-art results on ImageNet-C, ImageNet-P, and ImageNet-A.
Transfer learning is a cornerstone of computer vision, yet little work has been done to evaluate the relationship between architecture and transfer. An implicit hypothesis in modern computer vision research is that models that perform better on ImageNet necessarily perform better on other vision tasks. However, this hy…
Single neural network predicts ImageNet model parameters for faster training.
problem Training diverse ImageNet models requires significant resources and time.
method Trained a neural network to predict ImageNet model parameters and used them for initialization.
result Models initialized with predicted parameters converge faster and achieve competitive performance.
Study shows noisy data collection in ImageNet leads to biased model performance.
problem Noisy data collection in ImageNet datasets causes biased model performance.
method Human studies on ImageNet dataset to investigate design choices and their impact.
result Systematic misalignment between ImageNet benchmark and real-world task.
CNNs trained on ImageNet favor textures over shapes, but can learn shape-based recognition.
problem CNNs trained on ImageNet favor textures over shapes, leading to biased recognition.
method Evaluating CNNs and human observers on images with a texture-shape cue conflict.
result CNNs trained on ImageNet favor textures over shapes, but can learn shape-based recognition.
Improved ImageNet classification with semi-supervised learning.
problem Image classification with limited labeled data.
method Noisy Student Training: semi-supervised learning with noisy student models.
result 88.4% top-1 accuracy on ImageNet, 2.0% better than state-of-the-art.
New datasets make AI models fail reliably, highlighting weaknesses.
problem AI models' vulnerabilities to adversarial examples.
method Created challenging datasets with adversarial filtering, tested on various models.
result Existing models perform poorly on new datasets, revealing shared weaknesses.
New test sets show current image classifiers struggle with harder images.
problem Current image classifiers overfit to original test sets.
method Created new test sets for CIFAR-10 and ImageNet, evaluated various models.
result Accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet.
FID misaligns with human judgment due to reliance on ImageNet classes.
problem FID metric's reliance on ImageNet classes causes discrepancies with human evaluation.
method Investigated and visualized the feature space of FID and its relation to ImageNet classes.
result Aligning histograms of Top-N ImageNet classifications can reduce FID without improving quality. Diffusion models outperform GANs in image synthesis quality.
problem Improving image synthesis quality using generative models.
method Developed and optimized diffusion models, and introduced classifier guidance for conditional synthesis.
result Achieved superior image synthesis quality with diffusion models, matching or exceeding GANs in FID scores.
1-hidden layer networks lead to deep networks that outperform AlexNet on ImageNet.
problem Improving deep neural networks for image classification tasks.
method Sequentially building deep networks layer by layer using shallow networks as auxiliary problems.
result 11-layer network trained by this method exceeds several VGG models on ImageNet.
This paper benchmarks neural network robustness to corruptions and perturbations.
problem Establishing benchmarks for image classifier robustness to corruptions and perturbations.
method Developed ImageNet-C and ImageNet-P datasets to evaluate robustness to corruptions and perturbations, not adversarial attacks.
result There are negligible changes in relative corruption robustness from AlexNet to ResNet classifiers.
Improves synthetic-to-real generalization without real data.
problem Synthetic models struggle with real data generalization.
method Encourages similar ImageNet representations and automates learning rates.
result Significant improvement in synthetic-to-real generalization.
Certified robustness for ImageNet models with randomized smoothing.
problem Creating robust models against adversarial attacks.
method Randomized smoothing with Gaussian noise.
result Certified top-1 accuracy of 49% on ImageNet under small ℓ2 perturbations. Improved GAN training speed and quality with FastGAN.
problem Slower convergence and less expressive models in GAN training.
method Adversarial training with FastGAN algorithm.
result Better generation quality with less training time.
CLIP models robustness to spurious features is re-evaluated using a new dataset.
problem Existing robustness tests of CLIP models may not fully reflect their performance on spurious features.
method Crafted a new dataset (CounterAnimal) to reveal CLIP models' reliance on realistic spurious features.
result CLIP models are robust to spurious features learned from their training data, not ImageNet.
Prunes neural networks at initialization to save resources, achieving high accuracy.
problem Efficiently reducing resource requirements for neural networks at both training and test time.
method Gradient Signal Preservation (GraSP) to prune networks at initialization.
result Pruning 80% of VGG-16 weights on ImageNet with only a 1.6% drop in accuracy.
Improved self-supervised learning on ImageNet achieves top-1 accuracy of 77.1%.
problem Self-supervised ResNets underperform supervised learning on ImageNet.
method ReLICv2 combines explicit invariance loss with contrastive objective over varied data views.
result ReLICv2 achieves 77.1% top-1 accuracy on ImageNet, improving over previous state-of-the-art by 1.5%.
New framework evaluates model robustness against diverse, unforeseen attacks.
problem Real-world adversarial robustness is harder to assess than research suggests.
method ImageNet-UA framework for evaluating robustness against diverse, unseen attacks.
result Standard robustness measures fail to capture unforeseen robustness.
Biologically plausible learning algorithms can match BP on large datasets.
problem Learning algorithms that are biologically plausible often perform poorly on large datasets.
method Evaluation of sign-symmetry and feedback alignment algorithms on ImageNet and MS COCO.
result Sign-symmetry algorithm can match BP performance on ImageNet and MS COCO.
CEB enhances model resilience through simple entropy bottleneck.
problem Improving model robustness against adversarial attacks.
method Conditional Entropy Bottleneck (CEB) combined with data augmentation.
result CEB significantly boosts adversarial robustness on various benchmarks.
In this paper, we develop improved techniques for defending against adversarial examples at scale. First, we implement the state of the art version of adversarial training at unprecedented scale on ImageNet and investigate whether it remains effective in this setting - an important open scientific question (Athalye et …
New BNN architecture achieves high accuracy without complex tricks.
problem Training accurate Binary Neural Networks (BNNs) from scratch is challenging.
method Revisited design principles and a simple training strategy.
result BinaryDenseNet achieves high accuracy on ImageNet.
74.7 seconds to train ResNet-50 on ImageNet with 2,048 GPUs.
problem Training deep learning models on large mini-batches efficiently.
method Optimization methods applied to distributed deep learning.
result Top-1 validation accuracy of 75.08% with training time of 74.7 seconds.
ConvNets and Imagenet have driven the recent success of deep learning for image classification. However, the marked slowdown in performance improvement combined with the lack of robustness of neural networks to adversarial examples and their tendency to exhibit undesirable biases question the reliability of these metho…
NoisyMix boosts model robustness to common corruptions.
problem Improving robustness of neural networks in real-world applications.
method NoisyMix training scheme that uses noisy augmentations in input and feature space.
result NoisyMix produces more robust models with well-calibrated class membership probabilities.
Study reveals biases in ImageNet models are not sufficient for generalization.
problem Understanding and improving generalization of image classification models.
method Large-scale study of 48 ImageNet models trained via various methods.
result Biases identified in ImageNet models do not fully explain generalization.
In this brief technical report we introduce the CINIC-10 dataset as a plug-in extended alternative for CIFAR-10. It was compiled by combining CIFAR-10 with images selected and downsampled from the ImageNet database. We present the approach to compiling the dataset, illustrate the example images for different classes, g…
High error rates improve neural network performance and reduce power consumption.
problem Training neural networks with high Bit Error Rates (BERs).
method Trained three Binarized Convolutional Neural Network architectures on various datasets with high BERs.
result High BERs do not significantly degrade test accuracy, enabling more efficient hardware.
BagNet classifies images using local features, making decisions clearer and more understandable.
problem Understanding how deep neural networks make decisions is difficult.
method BagNet is a variant of ResNet-50 that classifies images based on local features without spatial ordering.
result BagNet achieves high accuracy on ImageNet (87.6% top-5 for 33 x 33 px features).
HMQ improves quantization for edge devices with mixed precision.
problem Efficient quantization for edge devices with uniform, power-of-two thresholds.
method Introduces HMQ, a mixed precision quantization block that repurposes Gumbel-Softmax for searching over quantization schemes.
result Achieves competitive and state-of-the-art results on ImageNet despite restrictions.
Improved disentanglement in VAEs using aggregated feature maps.
problem Improving disentanglement in Variational Autoencoders (VAEs).
method Regionally aggregated feature maps extracted from pre-trained CNNs on ImageNet.
result 2nd place in NeurIPS 2019 disentanglement challenge.
Study reveals statistical bias in dataset replication, reducing accuracy drop from 11-14% to 3.6%.
problem Statistical bias in dataset replication affects model generalization accuracy.
method Analyzed ImageNet-v2, identified and corrected for bias, and compared results.
result Correcting bias reduces accuracy drop from 11-14% to 3.6%.
DeformRS certifies deep networks against various input deformations.
problem Vulnerability of deep networks to input deformations.
method Randomized smoothing reformulation for general deformations.
result Certifies rich deformations including translations, rotations, scaling, and affine.
Enhances SSL methods with depth cues for better image understanding.
problem Lack of depth cues in 2D image pixel maps limits SSL performance.
method Integrates depth signals from a pretrained monocular RGB-to-depth model into contrastive learning frameworks.
result Improves SSL methods' robustness and generalization with depth signals.
Study shows image classifiers struggle with temporal changes in videos.
problem Temporal robustness of image classifiers in videos.
method Constructed datasets and evaluated various classifiers and detection models.
result Median classification accuracy drops of 16-10 points in robust datasets.
Free adversarial training improves robustness without generating adversarial examples.
problem Training robust models against adversarial attacks is costly and impractical for large-scale datasets.
method Recycles gradient information from parameter updates to generate adversarial examples.
result Free adversarial training achieves comparable robustness to PGD training at negligible cost.
ResNet models overfit benignly on Cifar10 but not on ImageNet due to label noise.
problem Understanding why benign overfitting fails in real-world classification tasks with label noise.
method Theoretical analysis of benign overfitting under a mild overparameterization setup.
result Benign overfitting can fail in the presence of label noise, unlike in heavy overparameterization settings.
Puzzle Mix optimizes mixup by leveraging saliency and local statistics.
problem Overfitting and adversarial vulnerability of deep networks.
method Puzzle Mix alternates between multi-label and saliency-based objectives for optimal mixing masks.
result Puzzle Mix outperforms other mixup methods in generalization and adversarial robustness.
This work increases shape bias in CNNs trained on ImageNet, improving robustness without accuracy gain.
problem Shape bias in CNNs trained on ImageNet.
method Uses domain-adversarial training to remove texture clues and increase shape bias.
result The method increases robustness of CNNs without improving accuracy.
New algorithm improves dataset distillation with 108% improvement on ImageNet.
problem Improving dataset distillation performance.
method Reparameterization and convexification of implicit gradients (RCIG).
result Establishes new state-of-the-art on various dataset distillation tasks.
Improved continuous-time consistency models for large-scale image generation.
problem Training instability and discretization errors in existing diffusion models.
method Unified theoretical framework, improved diffusion process, and network architecture.
result Trained continuous-time CMs at 1.5B parameters, achieving state-of-the-art FID scores.
Improved CNN training speed with K-FAC on large datasets.
problem Challenges in training large-scale neural networks with distributed processors.
method Scalable K-FAC design, layer-wise distribution, inverse-free second-order gradient evaluation, dynamic K-FAC update decoupling.
result Distributed K-FAC implementation converges to 75.9% MLPerf baseline in 18-25% less time than SGD.
Study shows current image classification models lack robustness to real-world dataset shifts.
problem Robustness of current image classification models to natural distribution shifts in real datasets.
method Evaluation of 204 ImageNet models in 213 different test conditions.
result Little to no transfer of robustness from synthetic to natural distribution shifts.
New metric detects adversarial samples with high accuracy.
problem Vulnerability of deep neural networks to adversarial samples.
method Analyzed adversarial samples through their contributions to principal components of images.
result Proposed new metric (k,p) point for measuring robustness to adversarial samples.
Baseline for few-shot image classification outperforms state-of-the-art.
problem Few-shot image classification challenges.
method Fine-tuning deep networks trained with cross-entropy loss, transductively.
result Outperforms state-of-the-art on various datasets.
Simple neural net outperforms complex uncertainty methods.
problem Reliable uncertainty estimation from deterministic models.
method A simple baseline using a single softmax neural net with residual connections and spectral normalization.
result Simple neural net outperforms DUQ and SNGP on uncertainty prediction.
Big models pretrain and fine-tune for semi-supervised learning on ImageNet.
problem Learning from few labeled examples with a large amount of unlabeled data.
method Unsupervised pretraining of a big ResNet model followed by supervised fine-tuning and distillation.
result 73.9% ImageNet top-1 accuracy with just 1% of labels (≤13 labeled images per class).