Modern deep neural network (DNN) trainings utilize various training techniques, e.g., nonlinear activation functions, batch normalization, skip-connections, etc. Despite their effectiveness, it is still mysterious how they help accelerate DNN trainings in practice. In this paper, we provide an empirical study of the regularization effect of these training techniques on DNN optimization. Specifically, we find that the optimization trajectories of successful DNN trainings consistently obey a certain regularity principle that regularizes the model update direction to be aligned with the trajectory direction. Theoretically, we show that such a regularity principle leads to a convergence guarantee in nonconvex optimization and the convergence rate depends on a regularization parameter. Empirically, we find that DNN trainings that apply the training techniques achieve a fast convergence and obey the regularity principle with a large regularization parameter, implying that the model updates are well aligned with the trajectory. On the other hand, DNN trainings without the training techniques have slow convergence and obey the regularity principle with a small regularization parameter, implying that the model updates are not well aligned with the trajectory. Therefore, different training techniques regularize the model update direction via the regularity principle to facilitate the convergence.

Reinforcement learning is a powerful framework for robots to acquire skills from experience, but often requires a substantial amount of online data collection. As a result, it is difficult to collect sufficiently diverse experiences that are needed for robots to generalize broadly. Videos of humans, on the other hand, are a readily available source of broad and interesting experiences. In this paper, we consider the question: can we perform reinforcement learning directly on experience collected by humans? This problem is particularly difficult, as such videos are not annotated with actions and exhibit substantial visual domain shift relative to the robot's embodiment. To address these challenges, we propose a framework for reinforcement learning with videos (RLV). RLV learns a policy and value function using experience collected by humans in combination with data collected by robots. In our experiments, we find that RLV is able to leverage such videos to learn challenging vision-based skills with less than half as many samples as RL methods that learn from scratch.

Abstract reasoning and logic inference are difficult problems for neural networks, yet essential to their applicability in highly structured domains. In this work we demonstrate that a well known technique such as spectral regularization can significantly boost the capabilities of a neural learner. We introduce the Neural Abstract Reasoner (NAR), a memory augmented architecture capable of learning and using abstract rules. We show that, when trained with spectral regularization, NAR achieves $78.8\%$ accuracy on the Abstraction and Reasoning Corpus, improving performance 4 times over the best known human hand-crafted symbolic solvers. We provide some intuition for the effects of spectral regularization in the domain of abstract reasoning based on theoretical generalization bounds and Solomonoff's theory of inductive inference.

Aiming at drastic speedup for point-feature embeddings at test time, we propose a new framework that uses a pair of multi-layer perceptrons (MLP) and a lookup table (LUT) to transform point-coordinate inputs into high-dimensional features. When compared with PointNet's feature embedding part realized by MLP that requires millions of dot products, the proposed framework at test time requires no such layers of matrix-vector products but requires only looking up the nearest entities from the tabulated MLP followed by interpolation, defined over discrete inputs on a 3D lattice that is substantially arranged irregularly. We call this framework LUTI-MLP: LUT Interpolation ML that provides a way to train end-to-end irregularly tabulated MLP coupled to a LUT in a specific manner without the need for any approximation at test time. LUTI-MLP also provides significant speedup for Jacobian computation of the embedding function wrt global pose coordinate on Lie algebra $\mathfrak{se}(3)$ at test time, which could be used for point-set registration problems. After extensive evaluation using the ModelNet40, we confirmed that the LUTI-MLP even with a small (e.g., $4^3$) lattice yields performance comparable to that of the MLP while achieving significant speedup: $100\times$ for the embedding, $12\times$ for the approximate Jacobian, and $860\times$ for the canonical Jacobian.

Modeling a structured, dynamic environment like a video game requires keeping track of the objects and their states (declarative knowledge) as well as predicting how objects behave (procedural knowledge). Black-box models with a monolithic hidden state often fail to apply procedural knowledge consistently and uniformly, i.e., they lack systematicity. For example, in a video game, correct prediction of one enemy's trajectory does not ensure correct prediction of another's. We address this issue via an architecture that factorizes declarative and procedural knowledge and that imposes modularity within each form of knowledge. The architecture consists of active modules called object files that maintain the state of a single object and invoke passive external knowledge sources called schemata that prescribe state updates. To use a video game as an illustration, two enemies of the same type will share schemata but will have separate object files to encode their distinct state (e.g., health, position). We propose to use attention to determine which object files to update, the selection of schemata, and the propagation of information between object files. The resulting architecture is a drop-in replacement conforming to the same input-output interface as normal recurrent networks (e.g., LSTM, GRU) yet achieves substantially better generalization on environments that have multiple object tokens of the same type, including a challenging intuitive physics benchmark.

Recently, deep learning approaches for MR motion artifact correction have been extensively studied. Although these approaches have shown high performance and reduced computational complexity compared to classical methods, most of them require supervised training using paired artifact-free and artifact-corrupted images, which may prohibit its use in many important clinical applications. For example, transient severe motion (TSM) due to acute transient dyspnea in Gd-EOB-DTPA-enhanced MR is difficult to control and model for paired data generation. To address this issue, here we propose a novel unsupervised deep learning scheme through outlier-rejecting bootstrap subsampling and aggregation. This is inspired by the observation that motions usually cause sparse k-space outliers in the phase encoding direction, so k-space subsampling along the phase encoding direction can remove some outliers and the aggregation step can further improve the results from the reconstruction network. Our method does not require any paired data because the training step only requires artifact-free images. Furthermore, to address the smoothing from potential bias to the artifact-free images, the network is trained in an unsupervised manner using optimal transport driven cycleGAN. We verify that our method can be applied for artifact correction from simulated motion as well as real motion from TSM successfully, outperforming existing state-of-the-art deep learning methods.

This paper is focused on the improvement the efficiency of the sparse convolutional neural networks (CNNs) layers on graphic processing units (GPU). The Nvidia deep neural network (cuDnn) library provides the most effective implementation of deep learning (DL) algorithms for GPUs. GPUs are one of the most efficient and commonly used accelerators for deep learning computations. The modern CNN models need megabytes of coefficients and needed millions MAC operations to perform convolution. One of the most common techniques for compressing CNN models is weight pruning. There are two main types of pruning: structural (based on removing whole weight channels) and non-structural (removing individual weights). The first enables much easier acceleration, but with this type it is difficult to achieve a sparsity level and accuracy as high as that obtained with the second type. Non-structural pruning with retraining can generate a matrix-weight up to $\sim90\%$ or more of sparsity in some deep CNN models. This work shows when is worth using a direct sparse operation to speed-up the calculation of the convolution layers. The VGG-16, CNN-non-static and 1x1 layers from ResNet models were used as a benchmarks. In addition, we present the impact of using reduced precision on time efficiency.

To theoretically quantify the generalization properties granted by adding the LRGA module to GNNs, we focus on a specific family of expressive GNNs and show that augmenting it with LRGA provides algorithmic alignment to a powerful graph isomorphism test, namely the 2-Folklore Weisfeiler-Lehman (2-FWL) algorithm. In more detail we: (i) consider the recent Random Graph Neural Network (RGNN) (Sato et al., 2020) framework and prove that it is universal in probability; (ii) show that RGNN augmented with LRGA aligns with 2-FWL update step via polynomial kernels; and (iii) bound the sample complexity of the kernel's feature map when learned with a randomly initialized two-layer MLP. From a practical point of view, augmenting existing GNN layers with LRGA produces state of the art results in current GNN benchmarks. Lastly, we observe that augmenting various GNN architectures with LRGA often closes the performance gap between different models. In many domains, data can be represented as a graph, where entities interact, have meaningful relations and a global structure. The need to be able to infer and gain a better understanding of such data rises in many instances such as social networks, citations and collaborations, chemoinformatics, epidemiology etc. In recent years, along with the major evolution of artificial neural networks, graph learning has also gained a new powerful tool - graph neural networks (GNNs). Since first originated (Gori et al., 2005; Scarselli et al., 2009) as recurrent algorithms, GNNs have become a central interest and the main tool in graph learning.

Sequence labeling (SL) is a fundamental research problem encompassing a variety of tasks, e.g., part-of-speech (POS) tagging, named entity recognition (NER), text chunking, etc. Though prevalent and effective in many downstream applications (e.g., information retrieval, question answering, and knowledge graph embedding), conventional sequence labeling approaches heavily rely on hand-crafted or language-specific features. Recently, deep learning has been employed for sequence labeling tasks due to its powerful capability in automatically learning complex features of instances and effectively yielding the stat-of-the-art performances. In this paper, we aim to present a comprehensive review of existing deep learning-based sequence labeling models, which consists of three related tasks, e.g., part-of-speech tagging, named entity recognition, and text chunking. Then, we systematically present the existing approaches base on a scientific taxonomy, as well as the widely-used experimental datasets and popularly-adopted evaluation metrics in the SL domain. Furthermore, we also present an in-depth analysis of different SL models on the factors that may affect the performance and future directions in the SL domain.

Object detection is an elementary Computer Vision task which deals with the classification of objects in a digital image to a particular class (such as airplanes, cars, humans, etc). This can further be used in the implementation of real-world systems like face detection, pedestrian detection, automated driving systems, video surveillance, among other applications. It has gathered a lot of attention in the last few years since it is closely related to video analysis. They also help provide keen insights on the image contents. In recent years, deep learning methods have gained momentum and are now able to learn large amount of features, at comparatively deeper levels, and are thus able to address the problems faced earlier in traditional network architectures, such as artificial neural networks.