Due to their hunger for big data, modern deep learning models are trained in parallel, often in distributed environments, where communication of model updates is the bottleneck. Various update compression (e.g., quantization, sparsification, dithering) techniques have been proposed in recent years as a successful tool to alleviate this problem. In this work, we introduce a new, remarkably simple and theoretically and practically effective compression technique, which we call natural compression (NC). Our technique is applied individually to all entries of the to-be-compressed update vector and works by randomized rounding to the nearest (negative or positive) power of two. NC is "natural" since the nearest power of two of a real expressed as a float can be obtained without any computation, simply by ignoring the mantissa. We show that compared to no compression, NC increases the second moment of the compressed vector by the tiny factor 9/8 only, which means that the effect of NC on the convergence speed of popular training algorithms, such as distributed SGD, is negligible. However, the communications savings enabled by NC are substantial, leading to 3-4x improvement in overall theoretical running time. For applications requiring more aggressive compression, we generalize NC to natural dithering, which we prove is exponentially better than the immensely popular random dithering technique. Our compression operators can be used on their own or in combination with existing operators for a more aggressive combined effect. Finally, we show that N is particularly effective for the in-network aggregation (INA) framework for distributed training, where the update aggregation is done on a switch, which can only perform integer computations.

We propose a novel unsupervised domain adaptation framework based on domain-specific batch normalization in deep neural networks. We aim to adapt to both domains by specializing batch normalization layers in convolutional neural networks while allowing them to share all other model parameters, which is realized by a two-stage algorithm. In the first stage, we estimate pseudo-labels for the examples in the target domain using an external unsupervised domain adaptation algorithm---for example, MSTN or CPUA---integrating the proposed domain-specific batch normalization. The second stage learns the final models using a multi-task classification loss for the source and target domains. Note that the two domains have separate batch normalization layers in both stages. Our framework can be easily incorporated into the domain adaptation techniques based on deep neural networks with batch normalization layers. We also present that our approach can be extended to the problem with multiple source domains. The proposed algorithm is evaluated on multiple benchmark datasets and achieves the state-of-the-art accuracy in the standard setting and the multi-source domain adaption scenario.

In machine learning, it is observed that probabilistic predictions sometimes disagree with averaged actual outcomes on certain subsets of data. This is also known as miscalibration that is responsible for unreliability and unfairness of practical machine learning systems. In this paper, we put forward an evaluation metric for calibration, coined field-level calibration error, that measures bias in predictions over the input fields that the decision maker concerns. We show that existing calibration methods perform poorly under our new metric. Specifically, after learning a calibration mapping over the validation dataset, existing methods have limited improvements in our error metric and completely fail to improve other non-calibration metrics such as the AUC score. We propose Neural Calibration, a new calibration method, which learns to calibrate by making full use of all input information over the validation set. We test our method on five large-scale real-world datasets. The results show that Neural Calibration significantly improves against uncalibrated predictions in all well-known metrics such as the negative log-likelihood, the Brier score, the AUC score, as well as our proposed field-level calibration error.

As one of the important functions of the intelligent transportation system (ITS), supply-demand prediction for autonomous vehicles provides a decision basis for its control. In this paper, we present two prediction models (i.e. ARLP model and Advanced ARLP model) based on two system environments that only the current day's historical data is available or several days' historical data are available. These two models jointly consider the spatial, temporal, and semantic relations. Spatial dependency is captured with residual network and dimension reduction. Short term temporal dependency is captured with LSTM. Long term temporal dependency and temporal shifting are captured with LSTM and attention mechanism. Semantic dependency is captured with multi-attention mechanism and autocorrelation coefficient method. Extensive experiments show that our frameworks provide more accurate and stable prediction results than the existing methods.

One may represent a graph using both its node-edge and its node-node incidence matrices. This choice of representation leads to two classes of equivariant neural networks for attributed graphs that we call incidence networks. Moving beyond graphs, incidence tensors can represent higher dimensional geometric data structures, such as attributed mesh and polytope. For example, a triangulated mesh can be represented using either a "homogeneous" node-node-node or an "inhomogeneous" node-edge-face incidence tensor. This is analogous to the choice of node-node vs. node-edge in graphs. We address the question of "which of these combinations of representation and deep model is more expressive?" and prove that for graphs, homogeneous and inhomogeneous models have equal expressive power. For higher dimensional incidence tensors, we prove that the inhomogeneous model, which is simpler and more practical, can also be more expressive. We demonstrate the effectiveness of incidence networks in quantum chemistry domain by reporting state-of-the-art on QM9 dataset, using both homogeneous and inhomogeneous representations.

State-of-the-art efficient model-based Reinforcement Learning (RL) algorithms typically act by iteratively solving empirical models, i.e., by performing \emph{full-planning} on Markov Decision Processes (MDPs) built by the gathered experience. In this paper, we focus on model-based RL in the finite-state finite-horizon MDP setting and establish that exploring with \emph{greedy policies} -- act by \emph{1-step planning} -- can achieve tight minimax performance in terms of regret, $\tilde{\mathcal{O}}(\sqrt{HSAT})$. Thus, full-planning in model-based RL can be avoided altogether without any performance degradation, and, by doing so, the computational complexity decreases by a factor of $S$. The results are based on a novel analysis of real-time dynamic programming, then extended to model-based RL. Specifically, we generalize existing algorithms that perform full-planning to such that act by 1-step planning. For these generalizations, we prove regret bounds with the same rate as their full-planning counterparts.

In this paper, we present a new dataset for Form Understanding in Noisy Scanned Documents (FUNSD). Form Understanding (FoUn) aims at extracting and structuring the textual content of forms. The dataset comprises 200 fully annotated real scanned forms. The documents are noisy and exhibit large variabilities in their representation making FoUn a challenging task. The proposed dataset can be used for various tasks including text detection, optical character recognition (OCR), spatial layout analysis and entity labeling/linking. To the best of our knowledge this is the first publicly available dataset with comprehensive annotations addressing the FoUn task. We also present a set of baselines and introduce metrics to evaluate performance on the FUNSD dataset. The FUNSD dataset can be downloaded at https://guillaumejaume.github. io/FUNSD/.

Autoencoders and their variations provide unsupervised models for learning low-dimensional representations for downstream tasks. Without proper regularization, autoencoder models are susceptible to the overfitting problem and the so-called posterior collapse phenomenon. In this paper, we introduce a quantization-based regularizer in the bottleneck stage of autoencoder models to learn meaningful latent representations. We combine both perspectives of Vector Quantized-Variational AutoEncoders (VQ-VAE) and classical denoising regularization schemes of neural networks. We interpret quantizers as regularizers that constrain latent representations while fostering a similarity mapping at the encoder. Before quantization, we impose noise on the latent variables and use a Bayesian estimator to optimize the quantizer-based representation. The introduced bottleneck Bayesian estimator outputs the posterior mean of the centroids to the decoder, and thus, is performing soft quantization of the latent variables. We show that our proposed regularization method results in improved latent representations for both supervised learning and clustering downstream tasks when compared to autoencoders using other bottleneck structures.

We consider learning ancestral causal relationships in high dimensions. Our approach is driven by a supervised learning perspective, with discrete indicators of causal relationships treated as labels to be learned from available data. We focus on the setting in which some causal (ancestral) relationships are known (via background knowledge or experimental data) and put forward a general approach that scales to large problems. This is motivated by problems in human biology which are characterized by high dimensionality and potentially many latent variables. We present a case study involving interventional data from human cells with total dimension $p \! \sim \! 19{,}000$. Performance is assessed empirically by testing model output against previously unseen interventional data. The proposed approach is highly effective and demonstrably scalable to the human genome-wide setting. We consider sensitivity to background knowledge and find that results are robust to nontrivial perturbations of the input information. We consider also the case, relevant to some applications, where the only prior information available concerns a small number of known ancestral relationships.

Deep neural networks (DNNs) have revolutionized AI due to their remarkable performance in pattern recognition, comprising of both memorizing complex training sets and demonstrating intelligence by generalizing to previously unseen data (test sets). The high generalization performance in DNNs has been explained by several mathematical tools, including optimization, information theory, and resilience analysis. In humans, it is the ability to abstract concepts from examples that facilitates generalization; this paper thus researches DNN generalization from that perspective. A recent computational neuroscience study revealed a correlation between abstraction and particular neural firing patterns. We express these brain patterns in a closed-form mathematical expression, termed the `Cognitive Neural Activation metric' (CNA) and apply it to DNNs. Our findings reveal parallels in the mechanism underlying abstraction in DNNs and those in the human brain. Beyond simply measuring similarity to human abstraction, the CNA is able to predict and rate how well a DNN will perform on test sets, and determines the best network architectures for a given task in a manner not possible with extant tools. These results were validated on a broad range of datasets (including ImageNet and random labeled datasets) and neural architectures.