While diffusion models have made remarkable progress in image generation, their outputs can still appear unrealistic and lack fine details, especially when using fewer number of neural function evaluations (NFEs) or lower guidance scales. To address this issue, we propose a novel momentum-based sampling technique, termed history-guided sampling (HiGS), which enhances quality and efficiency of diffusion sampling by integrating recent model predictions into each inference step. Specifically, HiGS leverages the difference between the current prediction and a weighted average of past predictions to steer the sampling process toward more realistic outputs with better details and structure. Our approach introduces practically no additional computation and integrates seamlessly into existing diffusion frameworks, requiring neither extra training nor fine-tuning. Extensive experiments show that HiGS consistently improves image quality across diverse models and architectures and under varying sampling budgets and guidance scales. Moreover, using a pretrained SiT model, HiGS achieves a new state-of-the-art FID of 1.61 for unguided ImageNet generation at 256$\times$256 with only 30 sampling steps (instead of the standard 250). We thus present HiGS as a plug-and-play enhancement to standard diffusion sampling that enables faster generation with higher fidelity.

Modern deep learning models in computer vision require large datasets of real images, which are difficult to curate and pose privacy and legal concerns, limiting their commercial use. Recent works suggest synthetic data as an alternative, yet models trained with it often underperform. This paper proposes a two-step approach to bridge this gap. First, we propose an improved neural fractal formulation through which we introduce a new class of synthetic data. Second, we propose reverse stylization, a technique that transfers visual features from a small, license-free set of real images onto synthetic datasets, enhancing their effectiveness. We analyze the domain gap between our synthetic datasets and real images using Kernel Inception Distance (KID) and show that our method achieves a significantly lower distributional gap compared to existing synthetic datasets. Furthermore, our experiments across different tasks demonstrate the practical impact of this reduced gap. We show that pretraining the EDM2 diffusion model on our synthetic dataset leads to an 11

We present a kernel-predicting neural denoising method for path-traced deep-Z images that facilitates their usage in animation and visual effects production. Deep-Z images provide enhanced flexibility during compositing as they contain color, opacity, and other rendered data at multiple depth-resolved bins within each pixel. However, they are subject to noise, and rendering until convergence is prohibitively expensive. The current state of the art in deep-Z denoising yields objectionable artifacts, and current neural denoising methods are incapable of handling the variable number of depth bins in deep-Z images. Our method extends kernel-predicting convolutional neural networks to address the challenges stemming from denoising deep-Z images. We propose a hybrid reconstruction architecture that combines the depth-resolved reconstruction at each bin with the flattened reconstruction at the pixel level. Moreover, we propose depth-aware neighbor indexing of the depth-resolved inputs to the convolution and denoising kernel application operators, which reduces artifacts caused by depth misalignment present in deep-Z images. We evaluate our method on a production-quality deep-Z dataset, demonstrating significant improvements in denoising quality and performance compared to the current state-of-the-art deep-Z denoiser. By addressing the significant challenge of the cost associated with rendering path-traced deep-Z images, we believe that our approach will pave the way for broader adoption of deep-Z workflows in future productions.

ReLU networks have remained the default choice for models in the area of image prediction despite their well-established spectral bias towards learning low frequencies faster, and consequently their difficulty of reproducing high frequency visual details. As an alternative, sin networks showed promising results in learning implicit representations of visual data. However training these networks in practically relevant settings proved to be difficult, requiring careful initialization, dealing with issues due to inconsistent gradients, and a degeneracy in local minima. In this work, we instead propose replacing a baseline network’s existing activations with a novel ensemble function with trainable parameters. The proposed MetaSin activation can be trained reliably without requiring intricate initialization schemes, and results in consistently lower test loss compared to alternatives. We demonstrate our method in the areas of Monte-Carlo denoising and image resampling where we set new state-of-the-art through a knowledge distillation based training procedure. We present ablations on hyper-parameter settings, comparisons with alternative activation function formulations, and discuss the use of our method in other domains, such as image classification.

Deep learning based methods for super-resolution have become state-of-the-art and outperform traditional approaches by a significant margin. From the initial models designed for fixed integer scaling factors (e.g. $\times 2$ or $\times 4)$ ), efforts were made to explore different directions such as modeling blur kernels or addressing non-integer scaling factors. However, existing works do not provide a sound framework to handle them jointly. In this paper we propose a framework for generic image resampling that not only addresses all the above mentioned issues but extends the sets of possible transforms from upscaling to generic transforms. A key aspect to unlock these capabilities is the faithful modeling of image warping and changes of the sampling rate during the training data preparation. This allows a localized representation of the implicit image degradation that takes into account the reconstruction kernel, the local geometric distortion and the anti-aliasing kernel. Using this spatially variant degradation map as conditioning for our resampling model, we can address with the same model both global transformations, such as upscaling or rotation, and locally varying transformations such lens distortion or undistortion. Another important contribution is the automatic estimation of the degradation map in this more complex resampling setting (i.e. blind image resampling). Fi-nally, we show that state-of-the-art results can be achieved by predicting kernels to apply on the input image instead of direct color prediction. This renders our model applicable for different types of data not seen during the training such as normals.

In this section we describe the network architectures in depth. Our methods architecture can be broken down into four distinct parts: degradation encoder, feature extractor, geometry encoder, and prediction network. The feature extractor is implemented using the ProSR architecture. The degradation encoder, geometry encoder, and prediction network are implemented as MLPs. A detailed breakdown is shown in Table 1. In total KARL c and KARL k consist of 2909212 and 2914866 parameters respectively.

We propose an adaptive sampling and reconstruction method for offline Monte Carlo rendering. Our method produces sampling maps constrained by a user-defined budget that minimize the expected future denoising error. Compared to other state-of-the-art methods, which produce the necessary training data on the fly by composing pre-rendered images, our method samples from analytic noise distributions instead. These distributions are compact and closely approximate the pixel value distributions stemming from Monte Carlo rendering. Our method can efficiently sample training data by leveraging only a few per-pixel statistics of the target distribution, which provides several benefits over the current state of the art. Most notably, our analytic distributions' modeling accuracy and sampling efficiency increase with sample count, essential for high-quality offline rendering. Although our distributions are approximate, our method supports joint end-to-end training of the sampling and denoising networks. Finally, we propose the addition of a global summary module to our architecture that accumulates valuable information from image regions outside of the network's receptive field. This information discourages sub-optimal decisions based on local information. Our evaluation against other state-of-the-art neural sampling methods demonstrates denoising quality and data efficiency improvements.

Although recent works have brought some insights into the performance improvement of techniques used in state-of-the-art deep-learning models, more work is needed to understand their generalization properties. We shed light on this matter by linking the loss function to the output's sensitivity to its input. We find a rather strong empirical relation between the output sensitivity and the variance in the bias-variance decomposition of the loss function, which hints on using sensitivity as a metric for comparing the generalization performance of networks, without requiring labeled data. We find that sensitivity is decreased by applying popular methods which improve the generalization performance of the model, such as (1) using a deep network rather than a wide one, (2) adding convolutional layers to baseline classifiers instead of adding fully-connected layers, (3) using batch normalization, dropout and max-pooling, and (4) applying parameter initialization techniques.

Knowledge graph embeddings rank among the most successful methods for link prediction in knowledge graphs, i.e., the task of completing an incomplete collection of relational facts. A downside of these models is their strong sensitivity to model hyperparameters, in particular regularizers, which have to be extensively tuned to reach good performance [Kadlec et al., 2017]. We propose an efficient method for large scale hyperparameter tuning by interpreting these models in a probabilistic framework. After a model augmentation that introduces per-entity hyperparameters, we use a variational expectation-maximization approach to tune thousands of such hyperparameters with minimal additional cost. Our approach is agnostic to details of the model and results in a new state of the art in link prediction on standard benchmark data.

Even though recent works have brought some insight into the performance improvement of techniques used in state-of-the-art deep-learning models, more work is needed to understand the generalization properties of over-parameterized deep neural networks. We shed light on this matter by linking the loss function to the output’s sensitivity to its input. We find a rather strong empirical relation between the output sensitivity and the variance in the bias-variance decomposition of the loss function, which hints on using sensitivity as a metric for comparing generalization performance of networks, without requiring labeled data. We find that sensitivity is decreased by applying popular methods which improve the generalization performance of the model, such as (1) using a deep network rather than a wide one, (2) adding convolutional layers to baseline classifiers instead of adding fully connected layers, (3) using batch normalization, dropout and max-pooling, and (4) applying parameter initialization techniques.

Personalization is pervasive in the online space as it leads to higher efficiency for the user and higher revenue for the platform by individualizing the most relevant content for each user. However, recent studies suggest that such personalization can learn and propagate systemic biases and polarize opinions; this has led to calls for regulatory mechanisms and algorithms that are constrained to combat bias and the resulting echo-chamber effect. We propose a versatile framework that allows for the possibility to reduce polarization in personalized systems by allowing the user to constrain the distribution from which content is selected. We then present a scalable algorithm with provable guarantees that satisfies the given constraints on the types of the content that can be displayed to a user, but- subject to these constraints- will continue to learn and personalize the content in order to maximize utility. We illustrate this framework on a curated dataset of online news articles that are conservative or liberal, show that it can control polarization, and examine the trade-off between decreasing polarization and the resulting loss to revenue. We further exhibit thefl exibility and scalability of our approach by framing the problem in terms of the more general diverse content selection problem and test it empirically on both a News dataset and the MovieLens dataset.

We develop a probabilistic extension of state-of-the-art embedding models for link prediction in relational knowledge graphs. Knowledge graphs are collections of relational facts, where each fact states that a certain relation holds between two entities, such as people, places, or objects. We argue that knowledge graphs should be treated within a Bayesian framework because even large knowledge graphs typically contain only few facts per entity, leading effectively to a small data problem where parameter uncertainty matters. We introduce a probabilistic reinterpretation of the DistMult (Yang et al., 2015) and ComplEx (Trouillon et al., 2016) models and employ variational inference to estimate a lower bound on the marginal likelihood of the data. We find that the main benefit of the Bayesian approach is that it allows for efficient, gradient based optimization over hyperparameters, which would lead to divergences in a non-Bayesian treatment. Models with such learned hyperparameters improve over the state-of-the-art by a significant margin, as we demonstrate on several benchmarks.

Personalization is pervasive in the online space as it leads to higher efficiency for the user and higher revenue for the platform by individualizing the most relevant content for each user. However, recent studies suggest that such personalization can learn and propagate systemic biases and polarize opinions; this has led to calls for regulatory mechanisms and algorithms that are constrained to combat bias and the resulting echo-chamber effect. We present our balanced news feed via a demo that displays a dashboard through which users can view the political leaning of their news consumption and set their polarization constraints. The balanced feed, as generated by the user-defined constraints, is then showcased side-by-side with the unconstrained (polarized) feed for comparison.

Learning the causal-interaction network of multivariate Hawkes processes is a useful task in many applications. Maximum-likelihood estimation is the most common approach to solve the problem in the presence of long observation sequences. However, when only short sequences are available, the lack of data amplifies the risk of overfitting and regularization becomes critical. Due to the challenges of hyper-parameter tuning, state-of-the-art methods only parameterize regularizers by a single shared hyper-parameter, hence limiting the power of representation of the model. To solve both issues, we develop in this work an efficient algorithm based on variational expectation-maximization. Our approach is able to optimize over an extended set of hyper-parameters. It is also able to take into account the uncertainty in the model parameters by learning a posterior distribution over them. Experimental results on both synthetic and real datasets show that our approach significantly outperforms state-of-the-art methods under short observation sequences.

Coordinate descent methods usually minimize a cost function by updating a random decision variable (corresponding to one coordinate) at a time. Ideally, we would update the decision variable that yields the largest decrease in the cost function. However, finding this coordinate would require checking all of them, which would effectively negate the improvement in computational tractability that coordinate descent is intended to afford. To address this, we propose a new adaptive method for selecting a coordinate. First, we find a lower bound on the amount the cost function decreases when a coordinate is updated. We then use a multi-armed bandit algorithm to learn which coordinates result in the largest lower bound by interleaving this learning with conventional coordinate descent updates except that the coordinate is selected proportionately to the expected decrease. We show that our approach improves the convergence of coordinate descent methods both theoretically and experimentally.

Personalization is pervasive in the online space as it leads to higher efficiency and revenue by allowing the most relevant content to be served to each user. However, recent studies suggest that personalization methods can propagate societal or systemic biases and polarize opinions; this has led to calls for regulatory mechanisms and algorithms to combat bias and inequality. Algorithmically, bandit optimization has enjoyed great success in learning user preferences and personalizing content or feeds accordingly. We propose an algorithmic framework that allows for the possibility to control bias or discrimination in such bandit-based personalization. Our model allows for the specification of general fairness constraints on the sensitive types of the content that can be displayed to a user. The challenge, however, is to come up with a scalable and low regret algorithm for the constrained optimization problem that arises. Our main technical contribution is a provably fast and low-regret algorithm for the fairness-constrained bandit optimization problem. Our proofs crucially leverage the special structure of our problem. Experiments on synthetic and real-world data sets show that our algorithmic framework can control bias with only a minor loss to revenue.

Many stochastic optimization algorithms work by estimating the gradient of the cost function on the fly by sampling datapoints uniformly at random from a training set. However, the estimator might have a large variance, which inadvertently slows down the convergence rate of the algorithms. One way to reduce this variance is to sample the datapoints from a carefully selected non-uniform distribution. In this work, we propose a novel non-uniform sampling approach that uses the multi-armed bandit framework. Theoretically, we show that our algorithm asymptotically approximates the optimal variance within a factor of 3. Empirically, we show that using this datapoint-selection technique results in a significant reduction in the convergence time and variance of several stochastic optimization algorithms such as SGD, SVRG and SAGA. This approach for sampling datapoints is general, and can be used in conjunction with any algorithm that uses an unbiased gradient estimation -- we expect it to have broad applicability beyond the specific examples explored in this work.

An individual's decisions are often guided by those of his or her peers, i.e., neighbors in a social network. Presumably, being privy to the experiences of others aids in learning and decision making, but how much advantage does an individual gain by observing her neighbors? Such problems make appearances in sociology and economics and, in this paper, we present a novel model to capture such decision-making processes and appeal to the classical multi-armed bandit framework to analyze it. Each individual, in addition to her own actions, can observe the actions and rewards obtained by her neighbors, and can use all of this information in order to minimize her own regret. We provide algorithms for this setting, both for stochastic and adversarial bandits, and show that their regret smoothly interpolates between the regret in the classical bandit setting and that of the full-information setting as a function of the neighbors' exploration. In the stochastic setting the additional information must simply be incorporated into the usual estimation of the rewards, while in the adversarial setting this is attained by constructing a new unbiased estimator for the rewards and appropriately bounding the amount of additional information provided by the neighbors. These algorithms are optimal up to log factors; despite the fact that the agents act independently and selfishly, this implies that it is an approximate Nash equilibria for all agents to use our algorithms. Further, we show via empirical simulations that our algorithms, often significantly, outperform existing algorithms that one could apply to this setting.