arXiv subjects

Alexander Keller

Publications and source records attributed to Alexander Keller.

48 records · Page 3Linked to original sources

arXiv2021

Massively Parallel Path Space Filtering

Restricting path tracing to a small number of paths per pixel for performance reasons rarely achieves a satisfactory image quality for scenes of interest. However, path space filtering may dramatically improve the visual quality by sharing information across vertices of paths classified as proximate. Unlike screen space-based approaches, these paths neither need to be present on the screen, nor is filtering restricted to the first intersection with the scene. While searching proximate vertices had been more expensive than filtering in screen space, we greatly improve over this performance penalty by storing, updating, and looking up the required information in a hash table. The keys are constructed from jittered and quantized information, such that only a single query very likely replaces costly neighborhood searches. A massively parallel implementation of the algorithm is demonstrated on a graphics processing unit (GPU).

cs.GR↗

arXiv2020

Neural Control Variates

We propose neural control variates (NCV) for unbiased variance reduction in parametric Monte Carlo integration. So far, the core challenge of applying the method of control variates has been finding a good approximation of the integrand that is cheap to integrate. We show that a set of neural networks can face that challenge: a normalizing flow that approximates the shape of the integrand and another neural network that infers the solution of the integral equation. We also propose to leverage a neural importance sampler to estimate the difference between the original integrand and the learned control variate. To optimize the resulting parametric estimator, we derive a theoretically optimal, variance-minimizing loss function, and propose an alternative, composite loss for stable online training in practice. When applied to light transport simulation, neural control variates are capable of matching the state-of-the-art performance of other unbiased approaches, while providing means to develop more performant, practical solutions. Specifically, we show that the learned light-field approximation is of sufficient quality for high-order bounces, allowing us to omit the error correction and thereby dramatically reduce the noise at the cost of negligible visible bias.

cs.LG↗

arXiv2020

Instant Quantization of Neural Networks using Monte Carlo Methods

Low bit-width integer weights and activations are very important for efficient inference, especially with respect to lower power consumption. We propose Monte Carlo methods to quantize the weights and activations of pre-trained neural networks without any re-training. By performing importance sampling we obtain quantized low bit-width integer values from full-precision weights and activations. The precision, sparsity, and complexity are easily configurable by the amount of sampling performed. Our approach, called Monte Carlo Quantization (MCQ), is linear in both time and space, with the resulting quantized, sparse networks showing minimal accuracy loss when compared to the original full-precision networks. Our method either outperforms or achieves competitive results on multiple benchmarks compared to previous quantization methods that do require additional training.

cs.LG↗

arXiv2019

Massively Parallel Construction of Radix Tree Forests for the Efficient Sampling of Discrete Probability Distributions

We compare different methods for sampling from discrete probability distributions and introduce a new algorithm which is especially efficient on massively parallel processors, such as GPUs. The scheme preserves the distribution properties of the input sequence, exposes constant time complexity on the average, and significantly lowers the average number of operations for certain distributions when sampling is performed in a parallel algorithm that requires synchronization afterwards. Avoiding load balancing issues of naïve approaches, a very efficient massively parallel construction algorithm for the required auxiliary data structure is complemented.

cs.DC↗

arXiv2019

Rethinking Full Connectivity in Recurrent Neural Networks

Recurrent neural networks (RNNs) are omnipresent in sequence modeling tasks. Practical models usually consist of several layers of hundreds or thousands of neurons which are fully connected. This places a heavy computational and memory burden on hardware, restricting adoption in practical low-cost and low-power devices. Compared to fully convolutional models, the costly sequential operation of RNNs severely hinders performance on parallel hardware. This paper challenges the convention of full connectivity in RNNs. We study structurally sparse RNNs, showing that they are well suited for acceleration on parallel hardware, with a greatly reduced cost of the recurrent operations as well as orders of magnitude less recurrent weights. Extensive experiments on challenging tasks ranging from language modeling and speech recognition to video action recognition reveal that structurally sparse RNNs achieve competitive performance as compared to fully-connected networks. This allows for using large sparse RNNs for a wide range of real-world tasks that previously were too costly with fully connected networks.

cs.LG↗

arXiv2019

Integral Equations and Machine Learning

As both light transport simulation and reinforcement learning are ruled by the same Fredholm integral equation of the second kind, reinforcement learning techniques may be used for photorealistic image synthesis: Efficiency may be dramatically improved by guiding light transport paths by an approximate solution of the integral equation that is learned during rendering. In the light of the recent advances in reinforcement learning for playing games, we investigate the representation of an approximate solution of an integral equation by artificial neural networks and derive a loss function for that purpose. The resulting Monte Carlo and quasi-Monte Carlo methods train neural networks with standard information instead of linear information and naturally are able to generate an arbitrary number of training samples. The methods are demonstrated for applications in light transport simulation.

cs.LG↗

arXiv2018

Fast, High Precision Ray/Fiber Intersection using Tight, Disjoint Bounding Volumes

Analyzing and identifying the shortcomings of current subdivision methods for finding intersections of rays with fibers defined by the surface of a circular contour swept along a Bézier curve, we present a new algorithm that improves precision and performance. Instead of the inefficient pruning using overlapping axis aligned bounding boxes and determining the closest point of approach of the ray and the curve, we prune using disjoint bounding volumes defined by cylinders and calculate the intersections on the limit surface. This in turn allows for computing accurate parametric position and normal in the point of intersection. The iteration requires only one bit per subdivision to avoid costly stack memory operations. At a low number of subdivisions, the performance of the high precision algorithm is competitive, while for a high number of subdivisions it dramatically outperforms the state-of-the-art. Besides an extensive mathematical analysis, source code is provided.

cs.GR↗

arXiv2018

Massively Parallel Stackless Ray Tracing of Catmull-Clark Subdivision Surfaces

We present a fast and efficient method for intersecting rays with Catmull-Clark subdivision surfaces. It takes advantage of the approximation democratized by OpenSubdiv, in which regular patches are represented by tensor product Bézier surfaces and irregular ones are approximated using Gregory patches. Our algorithm operates solely on the original patch data and can process both patch types simultaneously with only a small amount of control flow divergence. Besides introducing an optimized method to determine axis aligned bounding boxes of Gregory patches restricted in the parametric domain, several techniques are introduced that accelerate the recursive subdivision process including stackless operation, efficient work distribution, and control flow optimizations. The algorithm is especially useful for quick turnarounds during patch editing and animation playback.

cs.GR↗

arXiv2017

Learning Light Transport the Reinforced Way

We show that the equations of reinforcement learning and light transport simulation are related integral equations. Based on this correspondence, a scheme to learn importance while sampling path space is derived. The new approach is demonstrated in a consistent light transport simulation algorithm that uses reinforcement learning to progressively learn where light comes from. As using this information for importance sampling includes information about visibility, too, the number of light transport paths with zero contribution is dramatically reduced, resulting in much less noisy images within a fixed time budget.

cs.LG↗

arXiv2017

Comparison of predictions of the PC-SAFT equation of state and molecular simulations for the metastable region of binary mixtures

The metastable region of binary mixtures is examined with the PC-SAFT equa-tion of state (EOS) and compared to molecular simulations. The studied mixtures are Methane + Ethane, Methanol + Ethanol, Carbon Dioxide + Toluene, Carbon Dioxide + Hydrogen Chloride and Hydrogen Chloride + Toluene. In order to cal-culate the spinodal, the second partial derivatives with respect to the components molarities are analytically determined for the PC-SAFT EOS. Thermal properties as well as the spinodal points determined using PC-SAFT and molecular simula-tion are compared. For the studied metastable gas phases the results obtained with PC-SAFT and the molecular simulations agree well. For liquid phases the meta-stable region calculated with the PC-SAFT EOS is typically broader than that ob-tained from the molecular simulations. No unphysical results were obtained with PC-SAFT for metastable states. Hence, PC-SAFT EOS is found to be principally suitable for predicting metastable states and spinodals of mixtures.

physics.comp-ph↗

arXiv2017

Efficient Discovery of Ontology Functional Dependencies

Poor data quality has become a pervasive issue due to the increasing complexity and size of modern datasets. Constraint based data cleaning techniques rely on integrity constraints as a benchmark to identify and correct errors. Data values that do not satisfy the given set of constraints are flagged as dirty, and data updates are made to re-align the data and the constraints. However, many errors often require user input to resolve due to domain expertise defining specific terminology and relationships. For example, in pharmaceuticals, 'Advil' \emph{is-a} brand name for 'ibuprofen' that can be captured in a pharmaceutical ontology. While functional dependencies (FDs) have traditionally been used in existing data cleaning solutions to model syntactic equivalence, they are not able to model broader relationships (e.g., is-a) defined by an ontology. In this paper, we take a first step towards extending the set of data quality constraints used in data cleaning by defining and discovering \emph{Ontology Functional Dependencies} (OFDs). We lay out theoretical and practical foundations for OFDs, including a set of sound and complete axioms, and a linear inference procedure. We then develop effective algorithms for discovering OFDs, and a set of optimizations that efficiently prune the search space. Our experimental evaluation using real data show the scalability and accuracy of our algorithms.

cs.DB↗

arXiv2017

The Iray Light Transport Simulation and Rendering System

While ray tracing has become increasingly common and path tracing is well understood by now, a major challenge lies in crafting an easy-to-use and efficient system implementing these technologies. Following a purely physically-based paradigm while still allowing for artistic workflows, the Iray light transport simulation and rendering system allows for rendering complex scenes by the push of a button and thus makes accurate light transport simulation widely available. In this document we discuss the challenges and implementation choices that follow from our primary design decisions, demonstrating that such a rendering system can be made a practical, scalable, and efficient real-world application that has been adopted by various companies across many fields and is in use by many industry professionals today.

cs.GR↗