Abstract: Apparatuses, systems, and techniques to train neural networks. In at least one embodiment, a first normalization of learned parameters of one or more learned layers is performed during a forward pass of a training iteration and a second normalization of the learned parameters is performed during a parameter update phase of the training iteration. In at least one embodiment, the first normalization is performed using first scaling factors and the second normalization is performed using second scaling factors.

Abstract: Apparatuses, systems, and techniques to compute neural network parameters and to use a neural network to perform inference. In at least one embodiment, neural network parameters are computed, after training, by determining a weighted average of snapshots of averaged parameters that form a basis set of averaged parameter snapshots, each respective snapshot of averaged parameters including a plurality of network parameters averaged by a respective combination of an averaging function and one or more averaging parameters.

Abstract: Apparatuses, systems, and techniques to train neural networks and to use neural networks to perform inference. In at least one embodiment, a balanced concatenation layer performs a balanced concatenation operation during a forward pass of a training iteration during the training of a neural network. In at least one embodiment, a balanced concatenation layer performs a balanced concatenation operation during the use of a neural network to perform inference.

Abstract: A method and system are disclosed for training a model that implements a machine-learning algorithm. The technique utilizes latent descriptor vectors to change a multiple-valued output problem into a single-valued output problem and includes the steps of receiving a set of training data, processing, by a model, the set of training data to generate a set of output vectors, and adjusting a set of model parameters and component values for at least one latent descriptor vector in the plurality of latent descriptor vectors based on the set of output vectors. The set of training data includes a plurality of input vectors and a plurality of desired output vectors, and each input vector in the plurality of input vectors is associated with a particular latent descriptor vector in a plurality of latent descriptor vectors. Each latent descriptor vector comprises a plurality of scalar values that are initialized prior to training the model.

Abstract: User interfaces, methods and structures are described for intuitively and fluidly creating new artifacts from existing artifacts and for exploring latent spaces in a visual manner. In example embodiments, source artifacts are displayed along with a selector. The selector is operable to indicate a selected set of the source artifacts by establishing a selection region that includes portions of one or more of the source artifacts displayed. Source vectors are associated with the source artifacts in the selected set. One or more resultant vectors are determined based on the source vectors, and an output artifact is generated based on the one or more resultant vectors.

Abstract: Systems and methods are disclosed that improve output quality of any neural network, particularly an image generative neural network. In the real world, details of different scale tend to transform hierarchically. For example, moving a person's head causes the nose to move, which in turn moves the skin pores on the nose. Conventional generative neural networks do not synthesize images in a natural hierarchical manner: the coarse features seem to mainly control the presence of finer features, but not the precise positions of the finer features. Instead, much of the fine detail appears to be fixed to pixel coordinates which is a manifestation of aliasing. Aliasing breaks the illusion of a solid and coherent object moving in space. A generative neural network with reduced aliasing provides an architecture that exhibits a more natural transformation hierarchy, where the exact sub-pixel position of each feature is inherited from underlying coarse features.

Abstract: Systems and methods are disclosed for fused processing of a continuous mathematical operator. Fused processing of continuous mathematical operations, such as pointwise non-linear functions without storing intermediate results to memory improves performance when the memory bus bandwidth is limited. In an embodiment, a continuous mathematical operation including at least two of convolution, upsampling, pointwise non-linear function, and downsampling is executed to process input data and generate alias-free output data. In an embodiment, the input data is spatially tiled for processing in parallel such that the intermediate results generated during processing of the input data for each tile may be stored in a shared memory within the processor. Storing the intermediate data in the shared memory improves performance compared with storing the intermediate data to the external memory and loading the intermediate data from the external memory.

Abstract: A few-shot, unsupervised image-to-image translation (“FUNIT”) algorithm is disclosed that accepts as input images of previously-unseen target classes. These target classes are specified at inference time by only a few images, such as a single image or a pair of images, of an object of the target type. A FUNIT network can be trained using a data set containing images of many different object classes, in order to translate images from one class to another class by leveraging few input images of the target class. By learning to extract appearance patterns from the few input images for the translation task, the network learns a generalizable appearance pattern extractor that can be applied to images of unseen classes at translation time for a few-shot image-to-image translation task.

Abstract: Techniques are disclosed herein for generating a content item. The techniques include performing one or more first denoising operations based on an input and a first machine learning model to generate a first content item, and performing one or more second denoising operations based on the input, the first content item, and a second machine learning model to generate a second content item, where the first machine learning model is trained to denoise content items having an amount of corruption within a first corruption range, the second machine learning model is trained to denoise content items having an amount of corruption within a second corruption range, and the second corruption range is lower than the first corruption range.

Abstract: A method and system are disclosed for training a model that implements a machine-learning algorithm. The technique utilizes latent descriptor vectors to change a multiple-valued output problem into a single-valued output problem and includes the steps of receiving a set of training data, processing, by a model, the set of training data to generate a set of output vectors, and adjusting a set of model parameters and component values for at least one latent descriptor vector in the plurality of latent descriptor vectors based on the set of output vectors. The set of training data includes a plurality of input vectors and a plurality of desired output vectors, and each input vector in the plurality of input vectors is associated with a particular latent descriptor vector in a plurality of latent descriptor vectors. Each latent descriptor vector comprises a plurality of scalar values that are initialized prior to training the model.

Abstract: A method and system for performing novel image synthesis using generative networks are provided. The encoder-based model is trained to infer a 3D representation of an input image. A feature image is then generated using volume rendering techniques in accordance with the 3D representation. The feature image is then concatenated with a noisy image and processed by a denoiser network to predict an output image from a novel viewpoint that is consistent with the input image. The denoiser network can be a modified Noise Conditional Score Network (NCSN). In some embodiments, multiple input images or keyframes can be provided as input, and a different 3D representation is generated for each input image. The feature image is then generated, during volume rendering, by sampling each of the 3D representations and applying a mean-pooling operation to generate an aggregate feature image.

Abstract: User interfaces and methods are disclosed. In some embodiments, a plurality of source artifacts is displayed. A selector is operable to indicate a selected set of the source artifacts. An output artifact is displayed having an output attribute that represents a combination of source attributes from the source artifacts in the selected set. An amount of contribution to the first output attribute by respective ones of the source artifacts in the first selected set is based on a coordinate of the selector relative to coordinates of the source attributes in the first selected set.

Abstract: A style-based generative network architecture enables scale-specific control of synthesized output data, such as images. During training, the style-based generative neural network (generator neural network) includes a mapping network and a synthesis network. During prediction, the mapping network may be omitted, replicated, or evaluated several times. The synthesis network may be used to generate highly varied, high-quality output data with a wide variety of attributes. For example, when used to generate images of people's faces, the attributes that may vary are age, ethnicity, camera viewpoint, pose, face shape, eyeglasses, colors (eyes, hair, etc.), hair style, lighting, background, etc. Depending on the task, generated output data may include images, audio, video, three-dimensional (3D) objects, text, etc.

Abstract: In examples, a list of elements may be divided into spans and each span may be allocated a respective memory range for output based on a worst-case compression ratio of a compression algorithm that will be used to compress the span. Worker threads may output compressed versions of the spans to the memory ranges. To ensure placement constraints of a data structure will be satisfied, boundaries of the spans may be adjusted prior to compression. The size allocated to a span (e.g., each span) may be increased (or decreasing) to avoid padding blocks while allowing for the span's compressed data to use a block allocated to an adjacent span. Further aspects of the disclosure provide for compaction of the portions of compressed data in memory in order to free up space which may have been allocated to account for the memory gaps which may result from variable compression ratios.

Abstract: Techniques are disclosed herein for generating a content item. The techniques include receiving a content item and metadata indicating a level of corruption associated with the content item; and for each iteration included in a plurality of iterations: performing one or more operations to add corruption to a first version of the content item to generate a second version of the content item, and performing one or more operations to reduce corruption in the second version of the content item to generate a third version of the content item, wherein a level of corruption associated with the third version of the content item is less than a level of corruption associated with the first version of the content item.

Abstract: Techniques are disclosed herein for generating a content item. The techniques include receiving a content item and metadata indicating a level of corruption associated with the content item; and for each iteration included in a plurality of iterations: performing one or more operations to add corruption to a first version of the content item to generate a second version of the content item, and performing one or more operations to reduce corruption in the second version of the content item to generate a third version of the content item, wherein a level of corruption associated with the third version of the content item is less than a level of corruption associated with the first version of the content item.

Abstract: A three-dimensional (3D) density volume of an object is constructed from tomography images (e.g., x-ray images) of the object. The tomography images are projection images that capture all structures of an object (e.g., human body) between a beam source and imaging sensor. The beam effectively integrates along a path through the object producing a tomography image at the imaging sensor, where each pixel represents attenuation. A 3D reconstruction pipeline includes a first neural network model, a fixed function backprojection unit, and a second neural network model. Given information for the capture environment, the tomography images are processed by the reconstruction pipeline to produce a reconstructed 3D density volume of the object. In contrast with a set of 2D slices, the entire 3D density volume is reconstructed, so two-dimensional (2D) density images may be produced by slicing through any portion of the 3D density volume at any angle.

Abstract: A latent code defined in an input space is processed by the mapping neural network to produce an intermediate latent code defined in an intermediate latent space. The intermediate latent code may be used as appearance vector that is processed by the synthesis neural network to generate an image. The appearance vector is a compressed encoding of data, such as video frames including a person's face, audio, and other data. Captured images may be converted into appearance vectors at a local device and transmitted to a remote device using much less bandwidth compared with transmitting the captured images. A synthesis neural network at the remote device reconstructs the images for display.

Abstract: A generative adversarial neural network (GAN) learns a particular task by being shown many examples. In one scenario, a GAN may be trained to generate new images including specific objects, such as human faces, bicycles, etc. Rather than training a complex GAN having a predetermined topology of features and interconnections between the features to learn the task, the topology of the GAN is modified as the GAN is trained for the task. The topology of the GAN may be simple in the beginning and become more complex as the GAN learns during the training, eventually evolving to match the predetermined topology of the complex GAN. In the beginning the GAN learns large-scale details for the task (bicycles have two wheels) and later, as the GAN becomes more complex, learns smaller details (the wheels have spokes).

Abstract: A three-dimensional (3D) model of an object is recovered from two-dimensional (2D) images of the object. Each image in the set of 2D images includes the object captured from a different camera position and deformations of a base mesh that defines the 3D model may be computed corresponding to each image. The 3D model may also include a texture map that represents the lighting and material properties of the 3D model. Recovery of the 3D model relies on analytic antialiasing to provide a link between pixel colors in the 2D images and geometry of the 3D model. A modular differentiable renderer design yields high performance by leveraging existing, highly optimized hardware graphics pipelines to reconstruct the 3D model. The differential renderer renders images of the 3D model and differences between the rendered images and reference images are propagated backwards through the rendering pipeline to iteratively adjust the 3D model.