Abstract: A neural network model is trained for fairness and accuracy using both real and synthesized training data, such as images. During training a first sampling ratio between the real and synthesized training data is optimized. The first sampling ratio may comprise a value for each group (or attribute), where each value is optimized. A second sampling ratio defines relative amounts of training data that are used for each one of the groups. Furthermore, a neural network model accuracy and a fairness metric are both used for updating the first and second sampling ratios during training iterations. The neural network model may be trained using different classes of training data. The second sampling ratio may vary for each class.

Abstract: In various examples, systems and methods are disclosed relating to differentially private generative machine learning models. Systems and methods are disclosed for configuring generative models using privacy criteria, such as differential privacy criteria. The systems and methods can generate outputs representing content using machine learning models, such as diffusion models, that are determined in ways that satisfy differential privacy criteria. The machine learning models can be determined by diffusing the same training data to multiple noise levels.

Abstract: The disclosure provides a framework or system for learning visual representation using a large set of image/text pairs. The disclosure provides, for example, a method of visual representation learning, a joint representation learning system, and an artificial intelligence (AI) system that employs one or more of the trained models from the method or system. The AI system can be used, for example, in autonomous or semi-autonomous vehicles. In one example, the method of visual representation learning includes: (1) receiving a set of image embeddings from an image representation model and a set of text embeddings from a text representation model, and (2) training, employing mutual information, a critic function by learning relationships between the set of image embeddings and the set of text embeddings.

Abstract: Apparatuses, systems, and techniques are presented to remove unintended variations introduced into data. In at least one embodiment, a first image of an object can be generated based, at least in part, upon adding noise to, and removing the noise from, a second image of the object.

Abstract: Landmark detection refers to the detection of landmarks within an image or a video, and is used in many computer vision tasks such emotion recognition, face identity verification, hand tracking, gesture recognition, and eye gaze tracking. Current landmark detection methods rely on a cascaded computation through cascaded networks or an ensemble of multiple models, which starts with an initial guess of the landmarks and iteratively produces corrected landmarks which match the input more finely. However, the iterations required by current methods typically increase the training memory cost linearly, and do not have an obvious stopping criteria. Moreover, these methods tend to exhibit jitter in landmark detection results for video. The present disclosure improves current landmark detection methods by providing landmark detection using an iterative neural network.

Abstract: Approaches presented herein provide for the unconditional generation of novel three dimensional (3D) object shape representations, such as point clouds or meshes. In at least one embodiment, a first denoising diffusion model (DDM) can be trained to synthesize a 1D shape latent from Gaussian noise, and a second DDM can be trained to generate a set of latent points conditioned on this 1D shape latent. The shape latent and set of latent points can be provided to a decoder to generate a 3D point cloud representative of a random object from among the object classes on which the models were trained. A surface reconstruction process may be used to generate a surface mesh from this generated point cloud. Such an approach can scale to complex and/or multimodal distributions, and can be highly flexible as it can be adapted to various tasks such as multimodal voxel- or text-guided synthesis.

Abstract: Apparatuses, systems, and techniques to train a generative model based at least in part on a private dataset. In at least one embodiment, the generative model is trained based at least in part on a differentially private Sinkhorn algorithm, for example, using backpropagation with gradient descent to determine a gradient of a set of parameters of the generative models and modifying the set of parameters based at least in part on the gradient.

Abstract: Approaches presented herein provide for the generation of synthesized data from input noise using a denoising diffusion network. A higher order differential equation solver can be used for the denoising process, with one or more higher-order terms being distilled into one or more separate efficient neural networks. A separate, efficient neural network can be called together with a primary denoising model at inference time without significant loss in sampling efficiency. The separate neural network can provide information about the curvature (or other higher-order term) of the differential equation, representing a denoising trajectory, that can be used by the primary diffusion network to denoise the image using fewer denoising iterations.

Abstract: A machine learning model (MLM) may be trained and evaluated. Attribute-based performance metrics may be analyzed to identify attributes for which the MLM is performing below a threshold when each are present in a sample. A generative neural network (GNN) may be used to generate samples including compositions of the attributes, and the samples may be used to augment the data used to train the MLM. This may be repeated until one or more criteria are satisfied. In various examples, a temporal sequence of data items, such as frames of a video, may be generated which may form samples of the data set. Sets of attribute values may be determined based on one or more temporal scenarios to be represented in the data set, and one or more GNNs may be used to generate the sequence to depict information corresponding to the attribute values.

Abstract: One embodiment of the present invention sets forth a technique for executing a transformer neural network. The technique includes computing a first set of halting scores for a first set of tokens that has been input into a first layer of the transformer neural network. The technique also includes determining that a first halting score included in the first set of halting scores exceeds a threshold value. The technique further includes in response to the first halting score exceeding the threshold value, causing a first token that is included in the first set of tokens and is associated with the first halting score not to be processed by one or more layers within the transformer neural network that are subsequent to the first layer.

Abstract: The domain adaptation problem is addressed by using the predictions of a trained model over both source and target domain to retain the model with the assistance of an auxiliary model and a modified objective function. Inaccuracy in the model's predictions in the target domain is treated as noise and is reduced by using a robust learning framework during retraining, enabling unsupervised training in the target domain. Applications include object detection models, where noise in retraining is reduced by explicitly representing label noise and geometry noise in the objective function and using the ancillary model to inject information about label noise.

Abstract: Systems and methods described relate to the synthesis of content using generative models. In at least one embodiment, a score-based generative model can use a stochastic differential equation with critically-damped Langevin diffusion to learn to synthesize content. During a forward diffusion process, noise can be introduced into a set of auxiliary (e.g., “velocity”) values for an input image to learn a score function. This score function can be used with the stochastic differential equation during a reverse diffusion denoising process to remove noise from the image to generate a reconstructed version of the input image. A score matching objective for the critically-damped Langevin diffusion process can require only the conditional distribution learned from the velocity data. A stochastic differential equation-based integrator can then allow for efficient sampling from these critically-damped Langevin diffusion-based models.

Abstract: Apparatuses, systems, and techniques are presented to train and utilize one or more neural networks. A denoising diffusion generative adversarial network (denoising diffusion GAN) reduces a number of denoising steps during a reverse process. The denoising diffusion GAN does not assume a Gaussian distribution for large steps of the denoising process and applies a multi-model model to permit denoising with fewer steps. Systems and methods further minimize a divergence between a diffused real data distribution and a diffused generator distribution over several timesteps. Accordingly, various embodiments may enable faster sample generation, in which the samples are generated from noise using the denoising diffusion GAN.

Abstract: Apparatuses, systems, and techniques are presented to generate one or more images comprising one or more objects based, at least in part, on one or more dynamically configurable attributes of the one or objects. In at least one embodiment, one or more images comprising one or more objects can be generated based, at least in part, on one or more dynamically configurable attributes of the one or objects.

Abstract: One embodiment of the present invention sets forth a technique for training a generative model. The technique includes converting a first data point included in a training dataset into a first set of values associated with a base distribution for a score-based generative model. The technique also includes performing one or more denoising operations via the score-based generative model to convert the first set of values into a first set of latent variable values associated with a latent space. The technique further includes performing one or more additional operations to convert the first set of latent variable values into a second data point. Finally, the technique includes computing one or more losses based on the first data point and the second data point and generating a trained generative model based on the one or more losses, wherein the trained generative model includes the score-based generative model.

Abstract: Machine learning classification models which are robust against label noise are provided. Noise may be modelled explicitly by modelling “label flips”, where incorrect binary labels are “flipped” relative to their ground truth value. Distributions of label flips may be modelled as prior and posterior distributions in a flexible architecture for machine learning systems. An arbitrary classification model may be provided within the system. The classification model is made more robust to label noise by operation of the prior and posterior distributions. Particular prior and approximating posterior distributions are disclosed.

Abstract: One embodiment of the present invention sets forth a technique for generating data. The technique includes sampling from a first distribution associated with the score-based generative model to generate a first set of values. The technique also includes performing one or more denoising operations via the score-based generative model to convert the first set of values into a first set of latent variable values associated with a latent space. The technique further includes converting the first set of latent variable values into a generative output.

Abstract: Apparatuses, systems, and techniques to identify out-of-distribution input data in one or more neural networks. In at least one embodiment, a technique includes training one or more neural networks to infer a plurality of characteristics about input information based, at least in part, on the one or more neural networks being independently trained to infer each of the plurality of characteristics about the input information.

Abstract: Fully-supervised semantic segmentation machine learning models are augmented by ancillary machine learning models which generate high-detail predictions from low-detail, weakly-supervised data. The combined model can be trained over both fully- and weakly-supervised data. Only the primary model is required for inference, post-training. The combined model can be made self-correcting during training by adjusting the ancillary model's output based on parameters learned over both the fully- and weakly-supervised data. The self-correction module may combine the output of the primary and ancillary models in various ways, including through linear combinations and via neural networks. The self-correction module and ancillary model may benefit from disclosed pre-training techniques.

Abstract: Apparatuses, systems, and techniques to identify one or more images used to train one or more neural networks. In at least one embodiment, one or more images used to train one or more neural networks are identified, based on, for example, one or more labels of one or more objects within the one or more images.