A recently-proposed technique called self-adaptive training augments modern neural networks by allowing them to adjust training labels on the fly, to avoid overfitting to samples that may be mislabeled or otherwise non-representative. By combining the self-adaptive objective with mixup, we further improve the accuracy of self-adaptive models for image recognition; the resulting classifier obtains state-of-the-art accuracies on datasets corrupted with label noise. Robustness to label noise implies a lower generalization gap; thus, our approach also leads to improved generalizability. We find evidence that the Rademacher complexity of these algorithms is low, suggesting a new path towards provable generalization for this type of deep learning model. Last, we highlight a novel connection between difficulties accounting for rare classes and robustness under noise, as rare classes are in a sense indistinguishable from label noise. Our code can be found at https://github.com/Tuxianeer/generalizationconfusion.

We investigate whether the Wigner semi-circle and Marcenko-Pastur distributions, often used for deep neural network theoretical analysis, match empirically observed spectral densities. We find that even allowing for outliers, the observed spectral shapes strongly deviate from such theoretical predictions. This raises major questions about the usefulness of these models in deep learning. We further show that theoretical results, such as the layered nature of critical points, are strongly dependent on the use of the exact form of these limiting spectral densities. We consider two new classes of matrix ensembles; random Wigner/Wishart ensemble products and percolated Wigner/Wishart ensembles, both of which better match observed spectra. They also give large discrete spectral peaks at the origin, providing a theoretical explanation for the observation that various optima can be connected by one dimensional of low loss values. We further show that, in the case of a random matrix product, the weight of the discrete spectral component at $0$ depends on the ratio of the dimensions of the weight matrices.

We use a conditional deep convolutional generative adversarial network to predict the geopotential height of the 500 hPa pressure level, the two-meter temperature and the total precipitation for the next 24 hours over Europe. The proposed models are trained on 4 years of ERA5 reanalysis data from 2015-2018 with the goal to predict the associated meteorological fields in 2019. The forecasts show a good qualitative and quantitative agreement with the true reanalysis data for the geopotential height and two-meter temperature, while failing for total precipitation, thus indicating that weather forecasts based on data alone may be possible for specific meteorological parameters. We further use Monte-Carlo dropout to develop an ensemble weather prediction system based purely on deep learning strategies, which is computationally cheap and further improves the skill of the forecasting model, by allowing to quantify the uncertainty in the current weather forecast as learned by the model.

Many real-world situations allow for the acquisition of additional relevant information when making an assessment with limited or uncertain data. However, traditional ML approaches either require all features to be acquired beforehand or regard part of them as missing data that cannot be acquired. In this work, we propose models that dynamically acquire new features to further improve the prediction assessment. To trade off the improvement with the cost of acquisition, we leverage an information theoretic metric, conditional mutual information, to select the most informative feature to acquire. We leverage a generative model, arbitrary conditional flow (ACFlow), to learn the arbitrary conditional distributions required for estimating the information metric. We also learn a Bayesian network to accelerate the acquisition process. Our model demonstrates superior performance over baselines evaluated in multiple settings.

The accuracy and complexity of machine learning algorithms based on kernel optimization are limited by the set of kernels over which they are able to optimize. An ideal set of kernels should: admit a linear parameterization (for tractability); be dense in the set of all kernels (for robustness); be universal (for accuracy). The recently proposed Tesselated Kernels (TKs) is currently the only known class which meets all three criteria. However, previous algorithms for optimizing TKs were limited to classification and relied on Semidefinite Programming (SDP) - limiting them to relatively small datasets. By contrast, the 2-step algorithm proposed here scales to 10,000 data points and extends to the regression problem. Furthermore, when applied to benchmark data, the algorithm demonstrates significant improvement in performance over Neural Nets and SimpleMKL with similar computation time.

We develop a method to help quantify the impact different levels of mobility restrictions could have had on COVID-19 related deaths across nations. Synthetic control (SC) has emerged as a standard tool in such scenarios to produce counterfactual estimates if a particular intervention had not occurred, using just observational data. However, it remains an important open problem of how to extend SC to obtain counterfactual estimates if a particular intervention had occurred - this is exactly the question of the impact of mobility restrictions stated above. As our main contribution, we introduce synthetic interventions (SI), which helps resolve this open problem by allowing one to produce counterfactual estimates if there are multiple interventions of interest. We prove SI produces consistent counterfactual estimates under a tensor factor model. Our finite sample analysis shows the test error decays as $1/T_0$, where $T_0$ is the amount of observed pre-intervention data. As a special case, this improves upon the $1/\sqrt{T_0}$ bound on test error for SC in prior works. Our test error bound holds under a certain "subspace inclusion" condition; we furnish a data-driven hypothesis test with provable guarantees to check for this condition. This also provides a quantitative hypothesis test for when to use SC, currently absent in the literature. Technically, we establish the parameter estimation and test error for Principal Component Regression (a key subroutine in SI and several SC variants) under the setting of error-in-variable regression decays as $1/T_0$, where $T_0$ is the number of samples observed; this improves the best prior test error bound of $1/\sqrt{T_0}$. In addition to the COVID-19 case study, we show how SI can be used to run data-efficient, personalized randomized control trials using real data from a large e-commerce website and a large developmental economics study.

Latent position network models are a versatile tool in network science; applications include clustering entities, controlling for causal confounders, and defining priors over unobserved graphs. Estimating each node's latent position is typically framed as a Bayesian inference problem, with Metropolis within Gibbs being the most popular tool for approximating the posterior distribution. However, it is well-known that Metropolis within Gibbs is inefficient for large networks; the acceptance ratios are expensive to compute, and the resultant posterior draws are highly correlated. In this article, we propose an alternative Markov chain Monte Carlo strategy---defined using a combination of split Hamiltonian Monte Carlo and Firefly Monte Carlo---that leverages the posterior distribution's functional form for more efficient posterior computation. We demonstrate that these strategies outperform Metropolis within Gibbs and other algorithms on synthetic networks, as well as on real information-sharing networks of teachers and staff in a school district.

This paper extends the horseshoe prior of Carvalho et al. (2010) to the Bayesian quantile regression (HS-BQR) and provides a fast sampling algorithm that speeds up computation significantly in high dimensions. The performance of the HS-BQR is tested on large scale Monte Carlo simulations and an empirical application relevant to macroeoncomics. The Monte Carlo design considers several sparsity structures (sparse, dense, block) and error structures (i.i.d. errors and heteroskedastic errors). A number of LASSO based estimators (frequentist and Bayesian) are pitted against the HS-BQR to better gauge the performance of the method on the different designs. The HS-BQR yields just as good, or better performance than the other estimators considered when evaluated using coefficient bias and forecast error. We find that the HS-BQR is particularly potent in sparse designs and when estimating extreme quantiles. The simulations also highlight how the high dimensional quantile estimators fail to correctly identify the quantile function of the variables when both location and scale effects are present. In the empirical application, in which we evaluate forecast densities of US inflation, the HS-BQR provides well calibrated forecast densities whose individual quantiles, have the highest pseudo R squared, highlighting its potential for Value-at-Risk estimation.

We propose a framework for learning neural scene representations directly from images, without 3D supervision. Our key insight is that 3D structure can be imposed by ensuring that the learned representation transforms like a real 3D scene. Specifically, we introduce a loss which enforces equivariance of the scene representation with respect to 3D transformations. Our formulation allows us to infer and render scenes in real time while achieving comparable results to models requiring minutes for inference. In addition, we introduce two challenging new datasets for scene representation and neural rendering, including scenes with complex lighting and backgrounds. Through experiments, we show that our model achieves compelling results on these datasets as well as on standard ShapeNet benchmarks.

Transfer learning, which is to improve the learning performance in the target domain by leveraging useful knowledge from the source domain, often requires that those two domains are very close, which limits its application scope. Recently, distant transfer learning has been studied to transfer knowledge between two distant or even totally unrelated domains via auxiliary domains that are usually unlabeled as a bridge in the spirit of human transitive inference that it is possible to connect two completely unrelated concepts together through gradual knowledge transfer. In this paper, we study distant transfer learning by proposing a DeEp Random Walk basEd distaNt Transfer (DERWENT) method. Different from existing distant transfer learning models that implicitly identify the path of knowledge transfer between the source and target instances through auxiliary instances, the proposed DERWENT model can explicitly learn such paths via the deep random walk technique. Specifically, based on sequences identified by the random walk technique on a data graph where source and target data have no direct edges, the proposed DERWENT model enforces adjacent data points in a squence to be similar, makes the ending data point be represented by other data points in the same sequence, and considers weighted training losses of source data. Empirical studies on several benchmark datasets demonstrate that the proposed DERWENT algorithm yields the state-of-the-art performance.