In practice, by simply training a generator and a discriminator together consisting of multi-layer neural networks with non-linear activation functions, using local search algorithms such as stochastic gradient descent ascent (SGDA), the generator network can be trained efficiently to generate samples from highly-complicated distributions (such as the distribution of images). Despite the great empirical success of GAN, it remains to be one of the least understood models on the theory side of deep learning. Most of existing theories focus on the statistical properties of GANs at the global-optimum [15, 16, 20, 87]. However, on the training side, gradient descent ascent only enjoys efficient convergence to a global optimum when the loss function is convex-concave, or efficient convergence to a critical point in general settings [37, 38, 48, 53, 71, 73, 75, 77, 78]. Due to the extreme non-linearity of the networks in both the generator and the discriminator, it is highly unlikely that the training objective of GANs can be convex-concave. In particular, even if the generator and the discriminator are linear functions over prescribed feature mappings-- such as the neural tangent kernel (NTK) feature mappings [3, 8, 9, 17, 18, 32, 35, 40, 41, 47, 51, 54, 65, 69, 92, 97] -- the training objective can still be non-convex-concave.

We challenge a common assumption underlying most supervised deep learning: that a model makes a prediction depending only on its parameters and the features of a single input. To this end, we introduce a general-purpose deep learning architecture that takes as input the entire dataset instead of processing one datapoint at a time. Our approach uses self-attention to reason about relationships between datapoints explicitly, which can be seen as realizing non-parametric models using parametric attention mechanisms. However, unlike conventional non-parametric models, we let the model learn end-to-end from the data how to make use of other datapoints for prediction. Empirically, our models solve cross-datapoint lookup and complex reasoning tasks unsolvable by traditional deep learning models. We show highly competitive results on tabular data, early results on CIFAR-10, and give insight into how the model makes use of the interactions between points.

We introduce CAFLOW, a new diverse image-to-image translation model that simultaneously leverages the power of auto-regressive modeling and the modeling efficiency of conditional normalizing flows. We transform the conditioning image into a sequence of latent encodings using a multi-scale normalizing flow and repeat the process for the conditioned image. We model the conditional distribution of the latent encodings by modeling the auto-regressive distributions with an efficient multi-scale normalizing flow, where each conditioning factor affects image synthesis at its respective resolution scale. Our proposed framework performs well on a range of image-to-image translation tasks. It outperforms former designs of conditional flows because of its expressive auto-regressive structure.

Most studies on learning from noisy labels rely on unrealistic models of i.i.d. label noise, such as class-conditional transition matrices. More recent work on instance-dependent noise models are more realistic, but assume a single generative process for label noise across the entire dataset. We propose a more principled model of label noise that generalizes instance-dependent noise to multiple labelers, based on the observation that modern datasets are typically annotated using distributed crowdsourcing methods. Under our labeler-dependent model, label noise manifests itself under two modalities: natural error of good-faith labelers, and adversarial labels provided by malicious actors. We present two adversarial attack vectors that more accurately reflect the label noise that may be encountered in real-world settings, and demonstrate that under our multimodal noisy labels model, state-of-the-art approaches for learning from noisy labels are defeated by adversarial label attacks. Finally, we propose a multi-stage, labeler-aware, model-agnostic framework that reliably filters noisy labels by leveraging knowledge about which data partitions were labeled by which labeler, and show that our proposed framework remains robust even in the presence of extreme adversarial label noise.

Automatic speech recognition (ASR) models are typically designed to operate on a single input data type, e.g. a single or multi-channel audio streamed from a device. This design decision assumes the \textit{primary} input data source does not change and if an additional (\textit{auxiliary}) data source is occasionally available, it cannot be used. An ASR model that operates on both primary and auxiliary data can achieve better accuracy compared to a primary-only solution; and a model that can serve both \textit{primary-only} (PO) and \textit{primary-plus-auxiliary} (PPA) modes is highly desirable. In this work, we propose a unified ASR model that can serve both modes. We demonstrate its efficacy in a realistic scenario where a set of devices typically stream a single primary audio channel, and two additional auxiliary channels \textit{only when} upload bandwidth allows it. The architecture enables a unique methodology that uses both types of input audio during training time. Our proposed approach achieves up to 12.5\% relative word-error-rate reduction (WERR) compared to a PO baseline, and up to 16.0\% relative WERR in low-SNR conditions. The unique training methodology achieves up to 2.5\% relative WERR compared to a PPA baseline.

Computer-aided design (CAD) is the most widely used modeling approach for technical design. The typical starting point in these designs is 2D sketches which can later be extruded and combined to obtain complex three-dimensional assemblies. Such sketches are typically composed of parametric primitives, such as points, lines, and circular arcs, augmented with geometric constraints linking the primitives, such as coincidence, parallelism, or orthogonality. Sketches can be represented as graphs, with the primitives as nodes and the constraints as edges. Training a model to automatically generate CAD sketches can enable several novel workflows, but is challenging due to the complexity of the graphs and the heterogeneity of the primitives and constraints. In particular, each type of primitive and constraint may require a record of different size and parameter types. We propose SketchGen as a generative model based on a transformer architecture to address the heterogeneity problem by carefully designing a sequential language for the primitives and constraints that allows distinguishing between different primitive or constraint types and their parameters, while encouraging our model to re-use information across related parameters, encoding shared structure. A particular highlight of our work is the ability to produce primitives linked via constraints that enables the final output to be further regularized via a constraint solver. We evaluate our model by demonstrating constraint prediction for given sets of primitives and full sketch generation from scratch, showing that our approach significantly out performs the state-of-the-art in CAD sketch generation.

The advent of the transformer has sparked a quick growth in the size of language models, far outpacing hardware improvements. (Dense) transformers are expected to reach the trillion-parameter scale in the near future, for which training requires thousands or even tens of thousands of GPUs. We investigate the challenges of training at this scale and beyond on commercially available hardware. In particular, we analyse the shortest possible training time for different configurations of distributed training, leveraging empirical scaling laws for language models to estimate the optimal (critical) batch size. Contrary to popular belief, we find no evidence for a memory wall, and instead argue that the real limitation -- other than the cost -- lies in the training duration. In addition to this analysis, we introduce two new methods, \textit{layered gradient accumulation} and \textit{modular pipeline parallelism}, which together cut the shortest training time by half. The methods also reduce data movement, lowering the network requirement to a point where a fast InfiniBand connection is not necessary. This increased network efficiency also improve on the methods introduced with the ZeRO optimizer, reducing the memory usage to a tiny fraction of the available GPU memory.

Differential Privacy (DP) [13] is an important privacy-enhancing technology for private machine learning systems. It allows to measure and bound the risk associated with an individual participation in a computation. However, it was recently observed that DP learning systems may exacerbate bias and unfairness for different groups of individuals [3, 27, 22]. This paper builds on these important observations and sheds light on the causes of the disparate impacts arising in the problem of differentially private empirical risk minimization. It focuses on the accuracy disparity arising among groups of individuals in two well-studied DP learning methods: output perturbation [10] and differentially private stochastic gradient descent [2]. The paper analyzes which data and model properties are responsible for the disproportionate impacts, why these aspects are affecting different groups disproportionately, and proposes guidelines to mitigate these effects. The proposed approach is evaluated on several datasets and settings.

Aiming for a better integration of data-driven and linguistically-inspired approaches, we explore whether RST Nuclearity, assigning a binary assessment of importance between text segments, can be replaced by automatically generated, real-valued scores, in what we call a Weighted-RST framework. In particular, we find that weighted discourse trees from auxiliary tasks can benefit key NLP downstream applications, compared to nuclearity-centered approaches. We further show that real-valued importance distributions partially and interestingly align with the assessment and uncertainty of human annotators.

Inferring representations of 3D scenes from 2D observations is a fundamental problem of computer graphics, computer vision, and artificial intelligence. Emerging 3D-structured neural scene representations are a promising approach to 3D scene understanding. In this work, we propose a novel neural scene representation, Light Field Networks or LFNs, which represent both geometry and appearance of the underlying 3D scene in a 360-degree, four-dimensional light field parameterized via a neural implicit representation. Rendering a ray from an LFN requires only a *single* network evaluation, as opposed to hundreds of evaluations per ray for ray-marching or volumetric based renderers in 3D-structured neural scene representations. In the setting of simple scenes, we leverage meta-learning to learn a prior over LFNs that enables multi-view consistent light field reconstruction from as little as a single image observation. This results in dramatic reductions in time and memory complexity, and enables real-time rendering. The cost of storing a 360-degree light field via an LFN is two orders of magnitude lower than conventional methods such as the Lumigraph. Utilizing the analytical differentiability of neural implicit representations and a novel parameterization of light space, we further demonstrate the extraction of sparse depth maps from LFNs.