Neural Information Processing Systems http://nips.cc/

To train Variational Autoencoders (VAEs) to generate realistic imagery requires a loss function that reflects human perception of image similarity. We propose such a loss function based on Watson's perceptual model, which computes a weighted distance in frequency space and accounts for luminance and contrast masking. We extend the model to color images, increase its robustness to translation by using the Fourier Transform, remove artifacts due to splitting the image into blocks, and make it differentiable. In experiments, VAEs trained with the new loss function generated realistic, high-quality image samples. Compared to using the Euclidean distance and the Structural Similarity Index, the images were less blurry; compared to deep neural network based losses, the new approach required less computational resources and generated images with less artifacts.

The problem of inverting generative neural networks ( i.e., to recover the input latent

Conditional mean independence (CMI) testing is crucial for statistical tasks including model determination and variable importance evaluation. In this work, we introduce a novel population CMI measure and a bootstrap-based testing procedure that utilizes deep generative neural networks to estimate the conditional mean functions involved in the population measure. The test statistic is thoughtfully constructed to ensure that even slowly decaying nonparametric estimation errors do not affect the asymptotic accuracy of the test. Our approach demonstrates strong empirical performance in scenarios with high-dimensional covariates and response variable, can handle multivariate responses, and maintains nontrivial power against local alternatives outside an $n^{-1/2}$ neighborhood of the null hypothesis. We also use numerical simulations and real-world imaging data applications to highlight the efficacy and versatility of our testing procedure.

Weaknesses: I have one critical concern with this paper, which is that the proposed model presented here is extremely similar to one result from "A General and Adaptive Robust Loss Function", Jonathan T. Barron, CVPR, 2019. Section 3.1 of that paper (going from the arxiv version) has results on improving reconstruction/sampling quality from VAEs by using a loss on DCT coefficients of YUV images, very similar to what is done here. They also propose a loss with a heavy-tailed distribution that looks a lot like Equation 8 of this submission, and present a method where they optimize over the scale of the loss being imposed on each coefficient of the DCT (similar to this submission). And the improvement in sample/reconstruction quality they demonstrate looks a lot like what is shown in this submission. Given these overwhelming similarities, I'm unable to support the acceptance of this paper without a comparison to the approach presented in that work.

Three knowledgeable referees support acceptance, and I also recommend acceptance. The key contribution of this submission is a new reconstruction loss for VAEs (somewhat like JPEG loss) that matches human perception more closely than traditional VAE reconstruction losses (e.g. For applications where the goal is to generate sharp images rather than to maximize the likelihood of held-out data, the proposed method is a good alternative to other known ways of generating sharp images with VAEs (i.e, autoregressive/flow-based decoders and adversarial loss function). Unlike these alternatives, the proposed method introduces few additional parameters to learn from the data. R1's and R2's concern about the lack of quantitative measures of performance is justified, but the author response also makes a compelling point about the difficulty of picking a fair quantitative metric.

To train Variational Autoencoders (VAEs) to generate realistic imagery requires a loss function that reflects human perception of image similarity. We propose such a loss function based on Watson's perceptual model, which computes a weighted distance in frequency space and accounts for luminance and contrast masking. We extend the model to color images, increase its robustness to translation by using the Fourier Transform, remove artifacts due to splitting the image into blocks, and make it differentiable. In experiments, VAEs trained with the new loss function generated realistic, high-quality image samples. Compared to using the Euclidean distance and the Structural Similarity Index, the images were less blurry; compared to deep neural network based losses, the new approach required less computational resources and generated images with less artifacts.

The upcoming high-luminosity runs of the LHC will push the quantitative frontier of data taking to over 25-times its current rates. To ensure precision gains from such high statistics, this increase in experimental data needs to be met by an equal amount of simulation. The required computational power is predicted to outgrow the increase in budget in the coming years [1, 2]. One solution to this predicament is the augmentation of the expensive, Monte Carlo-based, simulation chain with generative machine learning. A special focus is often put on the costly detector simulation [3, 4]. This approach is only viable under the assumption that the generated data is not statistically limited to the size of the simulated training data. Previous studies have shown, for both toy data [5] and calorimeter images [6], that samples generated with generative neural networks can surpass the training statistics due to powerful interpolation abilities of the network in data space. These studies rely on comparing a distance measure between histograms of generated data and true hold-out data to the distance between smaller, statistically limited sets of Monte Carlo data and the hold-out set. The phenomenon of a generative model surpassing the precision of its training set is also known as amplification.

As the frequency of extreme weather events rises, it becomes increasingly crucial to understand and detect them at the earliest opportunity. Statistical models provide a way to enhance their interpretability and offer insights into the connections between extreme events. Since geophysical data is often coupled across both space and time this poses challenges for modeling, often leading to highly complex statistical models. For spatial data, such as precipitation, a common way to describe and analyze extremes are max-stable processes, which arise as the unique limit of pointwise maxima of random fields. These processes are an essential tool in analyzing spatial extremes (Davison et al., 2012), as they allow for flexible modeling of the underlying dependence structure. However, when it comes to modeling these extremes, usually only a few observations are available, even less so as the underlying process is usually changing across time. For that reason traditional statistical methods often fail to identify parameters correctly, particularly as these models are high dimensional and complex. Furthermore, estimating parameters becomes especially challenging when dealing with extreme values. Therefore, specifying a distribution rather than relying on point estimators can be beneficial for quantifying uncertainty.

This thesis investigates the application of state-of-the-art advances in generative neural networks for fast simulation of the Zero Degree Calorimeter (ZDC) neutron detector in the ALICE experiment at CERN. Traditional simulation methods using the GEANT Monte Carlo toolkit, while accurate, are computationally demanding. With increasing computational needs at CERN, efficient simulation techniques are essential. The thesis provides a comprehensive literature review on the application of neural networks in computer vision, fast simulations using machine learning, and generative neural networks in high-energy physics. The theory of the analyzed models is also discussed, along with technical aspects and the challenges associated with a practical implementation. The experiments evaluate various neural network architectures, including convolutional neural networks, vision transformers, and MLP-Mixers, as well as generative frameworks such as autoencoders, generative adversarial networks, vector quantization models, and diffusion models. Key contributions include the implementation and evaluation of these models, a significant improvement in the Wasserstein metric compared to existing methods with a low generation time of 5 milliseconds per sample, and the formulation of a list of recommendations for developing models for fast ZDC simulation. Open-source code and detailed hyperparameter settings are provided for reproducibility. Additionally, the thesis outlines future research directions to further enhance simulation fidelity and efficiency.