Pixel-space diffusion models outperform latent models on high-resolution image synthesis.
problem Efficiency and quality trade-off in high-resolution image synthesis.
method Sigmoid loss-weighting, simplified architecture, and resolution scaling.
result Achieved 1.5 FID on ImageNet512, new SOTA results on other datasets.
Latent diffusion improves robustness in missing data imputation.
problem Missing data imputation under MCAR corruption.
method Two-stage framework: VAE for latent feature learning, diffusion model in latent space.
result Latent diffusion maintains high quality and stability up to 50% missingness.
We introduce the BriarPatch, a pixel-space intervention that obscures sensitive attributes from representations encoded in pre-trained classifiers. The patches encourage internal model representations not to encode sensitive information, which has the effect of pushing downstream predictors towards exhibiting demograph…
New methods create counterfactuals for image regression models.
problem Creating interpretable explanations for regression models in images.
method Two methods using diffusion-based generative models to create counterfactuals.
result Diffusion-based methods produce realistic, semantic, and smooth counterfactuals.
Method generates joint posterior samples of source and foreground mass distributions for gravitational lensing.
problem Challenging inference problem for high-resolution, high signal-to-noise ratio gravitational lensing.
method Combines diffusion-based generative modeling and recurrent inference machines.
result Can model realistic gravitational lensing simulations down to the noise level.
This paper simplifies diffusion models for high resolution images.
problem Applying diffusion models to high resolution images is challenging.
method Adjust noise schedule, scale specific parts, add dropout, and use downsampling.
result Achieved state-of-the-art image generation performance.
Study fusion methods for financial image views to improve robustness against attacks.
problem Improving robustness of financial image views for next-day direction prediction.
method Same-source multi-view learning with early fusion and late fusion, using OHLCV and technical-indicator views, and evaluating pixel-space L-infinity attacks.
result Early fusion can suffer negative transfer under noisy settings, while late fusion is more reliable once labels stabilize.
Enhances OOD detection using latent diffusion for more robust and efficient training.
problem Improving reliability of machine learning models in real-world scenarios.
method Proposes Outlier-Aware Learning (OAL) framework that generates synthetic OOD data in latent space and uses MICL and KD modules.
result Demonstrates superior performance on benchmark datasets.
Comma.ai's approach to Artificial Intelligence for self-driving cars is based on an agent that learns to clone driver behaviors and plans maneuvers by simulating future events in the road. This paper illustrates one of our research approaches for driving simulation. One where we learn to simulate. Here we investigate v…
New method solves linear inverse problems using diffusion models.
problem Linear inverse problems in various domains.
method Posterior sampling with latent diffusion models.
result Provable sample recovery in linear models, outperforming previous methods.
Manipulating video content is easier than ever. Due to the misuse potential of manipulated content, multiple detection techniques that analyze the pixel data from the videos have been proposed. However, clever manipulators should also carefully forge the metadata and auxiliary header information, which is harder to do …
This paper takes a step towards temporal reasoning in a dynamically changing video, not in the pixel space that constitutes its frames, but in a latent space that describes the non-linear dynamics of the objects in its world. We introduce the Kalman variational auto-encoder, a framework for unsupervised learning of seq…
We propose a technique for making Convolutional Neural Network (CNN)-based models more transparent by visualizing input regions that are 'important' for predictions -- or visual explanations. Our approach, called Gradient-weighted Class Activation Mapping (Grad-CAM), uses class-specific gradient information to localize…
New reweighted losses improve diffusion model training and image quality.
problem Training and improving diffusion models for image generation.
method Constructing a cascade of time-dependent variational lower bounds.
result Significant improvements in pixel-space image modeling quality.
Improved sampling for Diffusion Models by accounting for covariance.
problem Sampling quality degradation in few-step Diffusion Models.
method Covariance-aware sampler using Tweedie's formula and Fourier-space decomposition.
result Consistently superior samples compared to state-of-the-art samplers.
We explore the use of Vector Quantized Variational AutoEncoder (VQ-VAE) models for large scale image generation. To this end, we scale and enhance the autoregressive priors used in VQ-VAE to generate synthetic samples of much higher coherence and fidelity than possible before. We use simple feed-forward encoder and dec…
A new method learns object representations from motion in slot representations.
problem Unsupervised object extraction from low-level visual data.
method Contrastive learning in slot representations, focusing on moving objects and distinct entities.
result Introduced a new evaluation metric to measure diversity of slot vectors.
Learning to predict future images from a video sequence involves the construction of an internal representation that models the image evolution accurately, and therefore, to some degree, its content and dynamics. This is why pixel-space video prediction may be viewed as a promising avenue for unsupervised feature learn…
A new method generates natural-looking adversarial examples by bounding internal activation values.
problem Creating natural-looking adversarial examples that closely mimic the original input.
method Bounding internal activation values through a distribution quantile bound and polynomial barrier loss function.
result Our attack achieves similar success and confidence levels as state-of-the-art methods but with more natural-looking perturbations.
TRIM improves interpretability of deep neural networks in cosmology.
problem Understanding which features a deep neural network uses in a transformed space.
method TRIM (Transformation IMportance) attributes importances to features in a transformed space.
result Combining TRIM with contextual decomposition helps identify physical features learned by DNNs.
WaveletGAN improves GANs by homogenizing noise through multi-channel wavelet filtering.
problem Current noise generation models in GANs struggle with homogenizing noise, leading to low-fidelity samples.
method Proposes a multi-channel wavelet-based filtering method to homogenize noise in GANs.
result WaveletGAN generates high-fidelity samples with the smallest FIDs on Fashion-MNIST, KMNIST, and SVHN datasets.
Automated skin lesion analysis is very crucial in clinical practice, as skin cancer is among the most common human malignancy. Existing approaches with deep learning have achieved remarkable performance on this challenging task, however, heavily relying on large-scale labelled datasets. In this paper, we present a nove…
Many machine learning image classifiers are vulnerable to adversarial attacks, inputs with perturbations designed to intentionally trigger misclassification. Current adversarial methods directly alter pixel colors and evaluate against pixel norm-balls: pixel perturbations smaller than a specified magnitude, according t…
Reflected Diffusion Models improve on score-based models by incorporating data constraints.
problem Numerical error in score-based models leads to unnatural samples.
method Reverses a reflected stochastic differential equation on data support, learning perturbed score function through generalized score matching loss.
result Improves sample quality and fidelity without architectural modifications.
Expectation maximization (EM) algorithm is to find maximum likelihood solution for models having latent variables. A typical example is Gaussian Mixture Model (GMM) which requires Gaussian assumption, however, natural images are highly non-Gaussian so that GMM cannot be applied to perform clustering task on pixel space…
Our attacks are stronger and faster under Wasserstein metric.
problem Vulnerability of deep models to adversarial attacks.
method Developed an exact yet efficient projection operator and used the Frank-Wolfe method.
result Generated much stronger attacks and improved model robustness.
Modern Hopfield networks help prevent forgetting in generative models after task changes.
problem How to prevent forgetting in generative models after task changes.
method Introduce intrinsic forgetting as an increase in Hopfield energy after task change, analyze memory replay effectiveness, and validate predictions in experiments.
result High-energy, outlier-like samples are more forgettable than cluster-like samples, and energy-based selection of replay samples mitigates forgetting.
PRoA assesses deep learning robustness against practical functional perturbations.
problem Inadequate practical robustness verification methods for deep learning systems.
method Probabilistic robustness assessment based on adaptive concentration.
result Statistical guarantees on probabilistic robustness against functional perturbations.
New findings show disentangled latent representations are not enough for robust compositional generalization.
problem Deep learning models struggle with compositional generalization, especially in out-of-distribution samples.
method Investigated a 2D Gaussian generation task with fully disentangled inputs, then forced disentangled latent representations into full-dimensional output space.
result Forcing disentangled latent representations into full-dimensional output space enables robust compositional generalization.
Paper introduces Latent-CLIP for efficient text-image comparison in latent space.
problem Efficiently compare text and images in latent space without costly decoding.
method Trains CLIP model in latent space, uses Latent-CLIP rewards for noise optimization, and guides generation away from harmful content.
result Latent-CLIP matches CLIP performance on text-image classification and harmful content detection.
A cost-effective method to generate high-resolution images using wavelet-based super-resolution.
problem High compute resources required for training state-of-the-art image generation models.
method Wavelet-based two-step training framework: low-frequency sampling followed by super-resolution.
result Achieves FID of 10.59 on ImageNet 512x512 with half the compute of BigGAN.
VAEBM combines VAEs and EBMs for efficient image generation.
problem Sampling from EBMs is slow and VAEs generate blurry images.
method VAEBM uses a VAE for quick sampling and an EBM to refine and exclude non-data regions.
result VAEBM generates high-quality images with short MCMC chains.
Improves GANs by sampling from an energy-based model induced by discriminator scores.
problem Improving the quality of images generated by GANs.
method DDLS (Discriminator Driven Latent Sampling) using the sum of latent prior log-density and discriminator output score.
result Significantly improves Inception Score on CIFAR-10 dataset.
Deep learning models generalize by extending decision boundaries outside the convex hull of training data.
problem Understanding how deep learning models generalize beyond their training data.
method Investigation of decision boundaries inside and outside the convex hull of training sets, using various neural network architectures and training regimes.
result Over-parameterization is necessary for deep learning models to extend decision boundaries outside the convex hull of their training data.
Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-res…
Few ideas have enjoyed as large an impact on deep learning as convolution. For any problem involving pixels or spatial representations, common intuition holds that convolutional neural networks may be appropriate. In this paper we show a striking counterexample to this intuition via the seemingly trivial coordinate tra…