Recent reinforcement learning algorithms, though achieving impressive results in various fields, suffer from brittle training effects such as regression in results and high sensitivity to initialization and parameters. We claim that some of the brittleness stems from variance differences, i.e. when different environment areas - states and/or actions - have different rewards variance. This causes two problems: First, the “Boring Areas Trap” in algorithms such as Q-learning, where moving between areas depends on the current area variance, and getting out of a boring area is hard due to its low variance. Second, the “Manipulative Consultant” problem, when value-estimation functions used in DQN and Actor-Critic algorithms influence the agent to prefer boring areas, regardless of the mean rewards return, as they maximize estimation precision rather than rewards. This sheds a new light on how exploration contribute to training, as it helps with both challenges. Cognitive experiments in humans showed that noised reward signals may paradoxically improve performance. We explain this using the two mentioned problems, claiming that both humans and algorithms may share similar challenges. Inspired by this result, we propose the Adaptive Symmetric Reward Noising (ASRN), by which we mean adding Gaussian noise to rewards according to their states’ estimated variance, thus avoiding the two problems while not affecting the environment’s mean rewards behavior. We conduct our experiments in a Multi Armed Bandit problem with variance differences. We demonstrate that a Q-learning algorithm shows the brittleness effect in this problem, and that the ASRN scheme can dramatically improve the results. We show that ASRN helps a DQN algorithm training process reach better results in an end to end autonomous driving task using the AirSim driving simulator.
http://arxiv.org/abs/1905.10144
In this paper, we describe features of two new robot prototypes that are actuated by an actively controlled gyro (flywheel, symmetric rotor) inside a hollow sphere that is located in the middle of the robots. No external actuators are used. The outside structure of the robots and the gyro are connected by a gimbal, which is similar in structure to a control moment gyroscope in spacecrafts. The joints of the gimbal can be actuated. In this way, the orientation of axis for the gyro in relation to the egg can be changed. Since the inertia of the fast rotating gyro is large in relation to the outside structure, a relative rotation of the axis against the outside structure results in a motion of the egg by inertia principle. In this way, we can use this principle for controlling the robot to move forward and turn around. The robots are shaped as spheroidal ellipsoids so they resemble eggs. So far, we have built and tested two robot prototypes.
http://arxiv.org/abs/1905.10134
Spoken dialogue systems that assist users to solve complex tasks such as movie ticket booking have become an emerging research topic in artificial intelligence and natural language processing areas. With a well-designed dialogue system as an intelligent personal assistant, people can accomplish certain tasks more easily via natural language interactions. Today there are several virtual intelligent assistants in the market; however, most systems only focus on single modality, such as textual or vocal interaction. A multimodal interface has various advantages: (1) allowing human to communicate with machines in a natural and concise form using the mixture of modalities that most precisely convey the intention to satisfy communication needs, and (2) providing more engaging experience by natural and human-like feedback. This paper explores a brand new research direction, which aims at bridging dialogue generation and facial expression synthesis for better multimodal interaction. The goal is to generate dialogue responses and simultaneously synthesize corresponding visual expressions on faces, which is also an ultimate step toward more human-like virtual assistants.
https://arxiv.org/abs/1905.11240
Search-based methods for hard combinatorial optimization are often guided by heuristics. Tuning heuristics in various conditions and situations is often time-consuming. In this paper, we propose NeuRewriter that learns a policy to pick heuristics and rewrite the local components of the current solution to iteratively improve it until convergence. The policy factorizes into a region-picking and a rule-picking component, each parameterized by a neural network trained with actor-critic methods in reinforcement learning. NeuRewriter captures the general structure of combinatorial problems and shows strong performance in three versatile tasks: expression simplification, online job scheduling and vehicle routing problems. NeuRewriter outperforms the expression simplification component in Z3; outperforms DeepRM and Google OR-tools in online job scheduling; and outperforms recent neural baselines and Google OR-tools in vehicle routing problems.
http://arxiv.org/abs/1810.00337
A robust and reliable semantic segmentation in adverse weather conditions is very important for autonomous cars, but most state-of-the-art approaches only achieve high accuracy rates in optimal weather conditions. The reason is that they are only optimized for good weather conditions and given noise models. However, most of them fail, if data with unknown disturbances occur, and their performance decrease enormously. One possibility to still obtain reliable results is to observe the environment with different sensor types, such as camera and lidar, and to fuse the sensor data by means of neural networks, since different sensors behave differently in diverse weather conditions. Hence, the sensors can complement each other by means of an appropriate sensor data fusion. Nevertheless, the fusion-based approaches are still susceptible to disturbances and fail to classify disturbed image areas correctly. This problem can be solved by means of a special training method, the so called Robust Learning Method (RLM), a method by which the neural network learns to handle unknown noise. In this work, two different sensor fusion architectures for semantic segmentation are compared and evaluated on several datasets. Furthermore, it is shown that the RLM increases the robustness in adverse weather conditions enormously, and achieve good results although no disturbance model has been learned by the neural network.
http://arxiv.org/abs/1905.10117
High-dimensional always-changing environments constitute a hard challenge for current reinforcement learning techniques. Artificial agents, nowadays, are often trained off-line in very static and controlled conditions in simulation such that training observations can be thought as sampled i.i.d. from the entire observations space. However, in real world settings, the environment is often non-stationary and subject to unpredictable, frequent changes. In this paper we propose and openly release CRLMaze, a new benchmark for learning continually through reinforcement in a complex 3D non-stationary task based on ViZDoom and subject to several environmental changes. Then, we introduce an end-to-end model-free continual reinforcement learning strategy showing competitive results with respect to four different baselines and not requiring any access to additional supervised signals, previously encountered environmental conditions or observations.
http://arxiv.org/abs/1905.10112
Drone racing is becoming a popular e-sport all over the world, and beating the best human drone race pilots has quickly become a new major challenge for artificial intelligence and robotics. In this paper, we propose a strategy for autonomous drone racing which is computationally more efficient than navigation methods like visual inertial odometry and simultaneous localization and mapping. This fast light-weight vision-based navigation algorithm estimates the position of the drone by fusing race gate detections with model dynamics predictions. Theoretical analysis and simulation results show the clear advantage compared to Kalman filtering when dealing with the relatively low frequency visual updates and occasional large outliers that occur in fast drone racing. Flight tests are performed on a tiny racing quadrotor named “Trashcan”, which was equipped with a Jevois smart-camera for a total of 72g. The test track consists of 3 laps around a 4-gate racing track. The gates spaced 4 meters apart and can be displaced from their supposed position. An average speed of 2m/s is achieved while the maximum speed is 2.6m/s. To the best of our knowledge, this flying platform is the smallest autonomous racing drone in the world and is 6 times lighter than the existing lightest autonomous racing drone setup (420g), while still being one of the fastest autonomous racing drones in the world.
http://arxiv.org/abs/1905.10110
Stereo is a prominent technique to infer dense depth maps from images, and deep learning further pushed forward the state-of-the-art, making end-to-end architectures unrivaled when enough data is available for training. However, deep networks suffer from significant drops in accuracy when dealing with new environments. Therefore, in this paper, we introduce Guided Stereo Matching, a novel paradigm leveraging a small amount of sparse, yet reliable depth measurements retrieved from an external source enabling to ameliorate this weakness. The additional sparse cues required by our method can be obtained with any strategy (e.g., a LiDAR) and used to enhance features linked to corresponding disparity hypotheses. Our formulation is general and fully differentiable, thus enabling to exploit the additional sparse inputs in pre-trained deep stereo networks as well as for training a new instance from scratch. Extensive experiments on three standard datasets and two state-of-the-art deep architectures show that even with a small set of sparse input cues, i) the proposed paradigm enables significant improvements to pre-trained networks. Moreover, ii) training from scratch notably increases accuracy and robustness to domain shifts. Finally, iii) it is suited and effective even with traditional stereo algorithms such as SGM.
http://arxiv.org/abs/1905.10107
Recently, fully convolutional neural networks (FCNs) have shown significant performance in image parsing, including scene parsing and object parsing. Different from generic object parsing tasks, hand parsing is more challenging due to small size, complex structure, heavy self-occlusion and ambiguous texture problems. In this paper, we propose a novel parsing framework, Multi-Scale Dual-Branch Fully Convolutional Network (MSDB-FCN), for hand parsing tasks. Our network employs a Dual-Branch architecture to extract features of hand area, paying attention on the hand itself. These features are used to generate multi-scale features with pyramid pooling strategy. In order to better encode multi-scale features, we design a Deconvolution and Bilinear Interpolation Block (DB-Block) for upsampling and merging the features of different scales. To address data imbalance, which is a common problem in many computer vision tasks as well as hand parsing tasks, we propose a generalization of Focal Loss, namely Multi-Class Balanced Focal Loss, to tackle data imbalance in multi-class classification. Extensive experiments on RHD-PARSING dataset demonstrate that our MSDB-FCN has achieved the state-of-the-art performance for hand parsing.
http://arxiv.org/abs/1905.10100
We propose a disentangled feature for weakly supervised multiclass sound event detection (SED), which helps ameliorate the performance and the training efficiency of class-wise attention based detection system by the introduction of more class-wise prior information as well as the network redundancy weight reduction. In this paper, we approach SED as a multiple instance learning (MIL) problem and utilize a neural network framework with class-wise attention pooling (cATP) module to solve it. Aiming at making finer detection even if there is only a small number of clips with less co-occurrence of the categories available in the training set, we optimize the high-level feature space of cATP-MIL by disentangling it based on class-wise identifiable information in the training set and obtain multiple different subspaces. Experiments show that our approach achieves competitive performance on Task4 of the DCASE2018 challenge.
http://arxiv.org/abs/1905.10091
The increasing interest in the usage of Artificial Intelligence techniques (AI) from the research community and industry to tackle “real world” problems, requires High Performance Computing (HPC) resources to efficiently compute and scale complex algorithms across thousands of nodes. Unfortunately, typical data scientists are not familiar with the unique requirements and characteristics of HPC environments. They usually develop their applications with high-level scripting languages or frameworks such as TensorFlow and the installation process often requires connection to external systems to download open source software during the build. HPC environments, on the other hand, are often based on closed source applications that incorporate parallel and distributed computing API’s such as MPI and OpenMP, while users have restricted administrator privileges, and face security restrictions such as not allowing access to external systems. In this paper we discuss the issues associated with the deployment of AI frameworks in a secure HPC environment and how we successfully deploy AI frameworks on SuperMUC-NG with Charliecloud.
http://arxiv.org/abs/1905.10090
Compared to RGB semantic segmentation, RGBD semantic segmentation can achieve better performance by taking depth information into consideration. However, it is still problematic for contemporary segmenters to effectively exploit RGBD information since the feature distributions of RGB and depth (D) images vary significantly in different scenes. In this paper, we propose an Attention Complementary Network (ACNet) that selectively gathers features from RGB and depth branches. The main contributions lie in the Attention Complementary Module (ACM) and the architecture with three parallel branches. More precisely, ACM is a channel attention-based module that extracts weighted features from RGB and depth branches. The architecture preserves the inference of the original RGB and depth branches, and enables the fusion branch at the same time. Based on the above structures, ACNet is capable of exploiting more high-quality features from different channels. We evaluate our model on SUN-RGBD and NYUDv2 datasets, and prove that our model outperforms state-of-the-art methods. In particular, a mIoU score of 48.3\% on NYUDv2 test set is achieved with ResNet50. We will release our source code based on PyTorch and the trained segmentation model at https://github.com/anheidelonghu/ACNet.
http://arxiv.org/abs/1905.10089
This study investigates the use of non-linear unsupervised dimensionality reduction techniques to compress a music dataset into a low-dimensional representation which can be used in turn for the synthesis of new sounds. We systematically compare (shallow) autoencoders (AEs), deep autoencoders (DAEs), recurrent autoencoders (with Long Short-Term Memory cells – LSTM-AEs) and variational autoencoders (VAEs) with principal component analysis (PCA) for representing the high-resolution short-term magnitude spectrum of a large and dense dataset of music notes into a lower-dimensional vector (and then convert it back to a magnitude spectrum used for sound resynthesis). Our experiments were conducted on the publicly available multi-instrument and multi-pitch database NSynth. Interestingly and contrary to the recent literature on image processing, we can show that PCA systematically outperforms shallow AE. Only deep and recurrent architectures (DAEs and LSTM-AEs) lead to a lower reconstruction error. The optimization criterion in VAEs being the sum of the reconstruction error and a regularization term, it naturally leads to a lower reconstruction accuracy than DAEs but we show that VAEs are still able to outperform PCA while providing a low-dimensional latent space with nice “usability” properties. We also provide corresponding objective measures of perceptual audio quality (PEMO-Q scores), which generally correlate well with the reconstruction error.
http://arxiv.org/abs/1806.04096
Crowd counting from a single image is a challenging task due to high appearance similarity, perspective changes and severe congestion. Many methods only focus on the local appearance features and they cannot handle the aforementioned challenges. In order to tackle them, we propose a Perspective Crowd Counting Network (PCC Net), which consists of three parts: 1) Density Map Estimation (DME) focuses on learning very local features for density map estimation; 2) Random High-level Density Classification (R-HDC) extracts global features to predict the coarse density labels of random patches in images; 3) Fore-/Background Segmentation (FBS) encodes mid-level features to segments the foreground and background. Besides, the DULR module is embedded in PCC Net to encode the perspective changes on four directions (Down, Up, Left and Right). The proposed PCC Net is verified on five mainstream datasets, which achieves the state-of-the-art performance on the one and attains the competitive results on the other four datasets. The source code is available at https://github.com/gjy3035/PCC-Net.
http://arxiv.org/abs/1905.10085
With the breakthroughs in deep learning, the recent years have witnessed a booming of artificial intelligence (AI) applications and services, spanning from personal assistant to recommendation systems to video/audio surveillance. More recently, with the proliferation of mobile computing and Internet-of-Things (IoT), billions of mobile and IoT devices are connected to the Internet, generating zillions Bytes of data at the network edge. Driving by this trend, there is an urgent need to push the AI frontiers to the network edge so as to fully unleash the potential of the edge big data. To meet this demand, edge computing, an emerging paradigm that pushes computing tasks and services from the network core to the network edge, has been widely recognized as a promising solution. The resulted new inter-discipline, edge AI or edge intelligence, is beginning to receive a tremendous amount of interest. However, research on edge intelligence is still in its infancy stage, and a dedicated venue for exchanging the recent advances of edge intelligence is highly desired by both the computer system and artificial intelligence communities. To this end, we conduct a comprehensive survey of the recent research efforts on edge intelligence. Specifically, we first review the background and motivation for artificial intelligence running at the network edge. We then provide an overview of the overarching architectures, frameworks and emerging key technologies for deep learning model towards training/inference at the network edge. Finally, we discuss future research opportunities on edge intelligence. We believe that this survey will elicit escalating attentions, stimulate fruitful discussions and inspire further research ideas on edge intelligence.
http://arxiv.org/abs/1905.10083
Web question answering (QA) has become an dispensable component in modern search systems, which can significantly improve users’ search experience by providing a direct answer to users’ information need. This could be achieved by applying machine reading comprehension (MRC) models over the retrieved passages to extract answers with respect to the search query. With the development of deep learning techniques, state-of-the-art MRC performances have been achieved by recent deep methods. However, existing studies on MRC seldom address the predictive uncertainty issue, i.e., how likely the prediction of an MRC model is wrong, leading to uncontrollable risks in real-world Web QA applications. In this work, we first conduct an in-depth investigation over the risk of Web QA. We then introduce a novel risk control framework, which consists of a qualify model for uncertainty estimation using the probe idea, and a decision model for selectively output. For evaluation, we introduce risk-related metrics, rather than the traditional EM and F1 in MRC, for the evaluation of risk-aware Web QA. The empirical results over both the real-world Web QA dataset and the academic MRC benchmark collection demonstrate the effectiveness of our approach.
http://arxiv.org/abs/1905.10077
It has been previously observed that training Variational Recurrent Autoencoders (VRAE) for text generation suffers from serious uninformative latent variables problem. The model would collapse into a plain language model that totally ignore the latent variables and can only generate repeating and dull samples. In this paper, we explore the reason behind this issue and propose an effective regularizer based approach to address it. The proposed method directly injects extra constraints on the posteriors of latent variables into the learning process of VRAE, which can flexibly and stably control the trade-off between the KL term and the reconstruction term, making the model learn dense and meaningful latent representations. The experimental results show that the proposed method outperforms several strong baselines and can make the model learn interpretable latent variables and generate diverse meaningful sentences. Furthermore, the proposed method can perform well without using other strategies, such as KL annealing.
http://arxiv.org/abs/1905.10072
Exploration bonuses derived from the novelty of observations in an environment have become a popular approach to motivate exploration for reinforcement learning (RL) agents in the past few years. Recent methods such as curiosity-driven exploration usually estimate the novelty of new observations by the prediction errors of their system dynamics models. In this paper, we introduce the concept of optical flow estimation from the field of computer vision to the RL domain and utilize the errors from optical flow estimation to evaluate the novelty of new observations. We introduce a flow-based intrinsic curiosity module (FICM) capable of learning the motion features and understanding the observations in a more comprehensive and efficient fashion. We evaluate our method and compare it with a number of baselines on several benchmark environments, including Atari games, Super Mario Bros., and ViZDoom. Our results show that the proposed method is superior to the baselines in certain environments, especially for those featuring sophisticated moving patterns or with high-dimensional observation spaces. We further analyze the hyper-parameters used in the training phase and discuss our insights into them.
http://arxiv.org/abs/1905.10071
Multi-step passenger demand forecasting is a crucial task in on-demand vehicle sharing services. However, predicting passenger demand over multiple time horizons is generally challenging due to the nonlinear and dynamic spatial-temporal dependencies. In this work, we propose to model multi-step citywide passenger demand prediction based on a graph and use a hierarchical graph convolutional structure to capture both spatial and temporal correlations simultaneously. Our model consists of three parts: 1) a long-term encoder to encode historical passenger demands; 2) a short-term encoder to derive the next-step prediction for generating multi-step prediction; 3) an attention-based output module to model the dynamic temporal and channel-wise information. Experiments on three real-world datasets show that our model consistently outperforms many baseline methods and state-of-the-art models.
http://arxiv.org/abs/1905.10069
Video object segmentation aims at accurately segmenting the target object regions across consecutive frames. It is technically challenging for coping with complicated factors (e.g., shape deformations, occlusion and out of the lens). Recent approaches have largely solved them by using backforth re-identification and bi-directional mask propagation. However, their methods are extremely slow and only support offline inference, which in principle cannot be applied in real time. Motivated by this observation, we propose a new detection-based paradigm for video object segmentation. We propose an unified One-Pass Video Segmentation framework (OVS-Net) for modeling spatial-temporal representation in an end-to-end pipeline, which seamlessly integrates object detection, object segmentation, and object re-identification. The proposed framework lends itself to one-pass inference that effectively and efficiently performs video object segmentation. Moreover, we propose a mask guided attention module for modeling the multi-scale object boundary and multi-level feature fusion. Experiments on the challenging DAVIS 2017 demonstrate the effectiveness of the proposed framework with comparable performance to the state-of-the-art, and the great efficiency about 11.5 fps towards pioneering real-time work to our knowledge, more than 5 times faster than other state-of-the-art methods.
http://arxiv.org/abs/1905.10064
Unsupervised text style transfer aims to transfer the underlying style of text but keep its main content unchanged without parallel data. Most existing methods typically follow two steps: first separating the content from the original style, and then fusing the content with the desired style. However, the separation in the first step is challenging because the content and style interact in subtle ways in natural language. Therefore, in this paper, we propose a dual reinforcement learning framework to directly transfer the style of the text via a one-step mapping model, without any separation of content and style. Specifically, we consider the learning of the source-to-target and target-to-source mappings as a dual task, and two rewards are designed based on such a dual structure to reflect the style accuracy and content preservation, respectively. In this way, the two one-step mapping models can be trained via reinforcement learning, without any use of parallel data. Automatic evaluations show that our model outperforms the state-of-the-art systems by a large margin, especially with more than 8 BLEU points improvement averaged on two benchmark datasets. Human evaluations also validate the effectiveness of our model in terms of style accuracy, content preservation and fluency. Our code and data, including outputs of all baselines and our model are available at https://github.com/luofuli/DualLanST.
http://arxiv.org/abs/1905.10060
Multi-view facial expression recognition (FER) is a challenging task because the appearance of an expression varies in poses. To alleviate the influences of poses, recent methods either perform pose normalization or learn separate FER classifiers for each pose. However, these methods usually have two stages and rely on good performance of pose estimators. Different from existing methods, we propose a pose-adaptive hierarchical attention network (PhaNet) that can jointly recognize the facial expressions and poses in unconstrained environment. Specifically, PhaNet discovers the most relevant regions to the facial expression by an attention mechanism in hierarchical scales, and the most informative scales are then selected to learn the pose-invariant and expression-discriminative representations. PhaNet is end-to-end trainable by minimizing the hierarchical attention losses, the FER loss and pose loss with dynamically learned loss weights. We validate the effectiveness of the proposed PhaNet on three multi-view datasets (BU-3DFE, Multi-pie, and KDEF) and two in-the-wild FER datasets (AffectNet and SFEW). Extensive experiments demonstrate that our framework outperforms the state-of-the-arts under both within-dataset and cross-dataset settings, achieving the average accuracies of 84.92\%, 93.53\%, 88.5\%, 54.82\% and 31.25\% respectively.
http://arxiv.org/abs/1905.10059
Algorithms for clustering points in metric spaces is a long-studied area of research. Clustering has seen a multitude of work both theoretically, in understanding the approximation guarantees possible for many objective functions such as k-median and k-means clustering, and experimentally, in finding the fastest algorithms and seeding procedures for Lloyd’s algorithm. The performance of a given clustering algorithm depends on the specific application at hand, and this may not be known up front. For example, a “typical instance” may vary depending on the application, and different clustering heuristics perform differently depending on the instance. In this paper, we define an infinite family of algorithms generalizing Lloyd’s algorithm, with one parameter controlling the initialization procedure, and another parameter controlling the local search procedure. This family of algorithms includes the celebrated k-means++ algorithm, as well as the classic farthest-first traversal algorithm. We design efficient learning algorithms which receive samples from an application-specific distribution over clustering instances and learn a near-optimal clustering algorithm from the class. We show the best parameters vary significantly across datasets such as MNIST, CIFAR, and mixtures of Gaussians. Our learned algorithms never perform worse than k-means++, and on some datasets we see significant improvements.
http://arxiv.org/abs/1809.06987
Manifold amount of video data gets generated every minute as we read this document, ranging from surveillance to broadcasting purposes. There are two roadblocks that restrain us from using this data as such, first being the storage which restricts us from only storing the information based on the hardware constraints. Secondly, the computation required to process this data is highly expensive which makes it infeasible to work on them. Compressive sensing(CS)[2] is a signal process technique[11], through optimization, the sparsity of a signal can be exploited to recover it from far fewer samples than required by the Shannon-Nyquist sampling theorem. There are two conditions under which recovery is possible. The first one is sparsity which requires the signal to be sparse in some domain. The second one is incoherence which is applied through the isometric property which is sufficient for sparse signals[9][10]. To sustain these characteristics, preserving all attributes in the uncompressed domain would help any kind of in this field. However, existing dataset fallback in terms of continuous tracking of all the object present in the scene, very few video datasets have comprehensive continuous tracking of objects. To address these problems collectively, in this work we propose a new comprehensive video dataset, where the data is compressed using pixel-wise coded exposure [3] that resolves various other impediments.
http://arxiv.org/abs/1905.10054
An effective and efficient person re-identification (ReID) algorithm will alleviate painful video watching, and accelerate the investigation progress. Recently, with the explosive requirements of practical applications, a lot of research efforts have been dedicated to heterogeneous person re-identification (He-ReID). In this paper, we review the state-of-the-art methods comprehensively with respect to four main application scenarios – low-resolution, infrared, sketch and text. We begin with a comparison between He-ReID and the general Homogeneous ReID (Ho-ReID) task. Then, we survey the models that have been widely employed in He-ReID. Available existing datasets for performing evaluation are briefly described. We then summarize and compare the representative approaches. Finally, we discuss some future research directions.
http://arxiv.org/abs/1905.10048
In this paper we study yes/no questions that are naturally occurring — meaning that they are generated in unprompted and unconstrained settings. We build a reading comprehension dataset, BoolQ, of such questions, and show that they are unexpectedly challenging. They often query for complex, non-factoid information, and require difficult entailment-like inference to solve. We also explore the effectiveness of a range of transfer learning baselines. We find that transferring from entailment data is more effective than transferring from paraphrase or extractive QA data, and that it, surprisingly, continues to be very beneficial even when starting from massive pre-trained language models such as BERT. Our best method trains BERT on MultiNLI and then re-trains it on our train set. It achieves 80.4% accuracy compared to 90% accuracy of human annotators (and 62% majority-baseline), leaving a significant gap for future work.
http://arxiv.org/abs/1905.10044
Generative models often use human evaluations to measure the perceived quality of their outputs. Automated metrics are noisy indirect proxies, because they rely on heuristics or pretrained embeddings. However, up until now, direct human evaluation strategies have been ad-hoc, neither standardized nor validated. Our work establishes a gold standard human benchmark for generative realism. We construct Human eYe Perceptual Evaluation (HYPE) a human benchmark that is (1) grounded in psychophysics research in perception, (2) reliable across different sets of randomly sampled outputs from a model, (3) able to produce separable model performances, and (4) efficient in cost and time. We introduce two variants: one that measures visual perception under adaptive time constraints to determine the threshold at which a model’s outputs appear real (e.g. 250ms), and the other a less expensive variant that measures human error rate on fake and real images sans time constraints. We test HYPE across six state-of-the-art generative adversarial networks and two sampling techniques on conditional and unconditional image generation using four datasets: CelebA, FFHQ, CIFAR-10, and ImageNet. We find that HYPE can track model improvements across training epochs, and we confirm via bootstrap sampling that HYPE rankings are consistent and replicable.
http://arxiv.org/abs/1904.01121
In this paper, we introduce and tackle the Outline Generation (OG) task, which aims to unveil the inherent content structure of a multi-paragraph document by identifying its potential sections and generating the corresponding section headings. Without loss of generality, the OG task can be viewed as a novel structured summarization task. To generate a sound outline, an ideal OG model should be able to capture three levels of coherence, namely the coherence between context paragraphs, that between a section and its heading, and that between context headings. The first one is the foundation for section identification, while the latter two are critical for consistent heading generation. In this work, we formulate the OG task as a hierarchical structured prediction problem, i.e., to first predict a sequence of section boundaries and then a sequence of section headings accordingly. We propose a novel hierarchical structured neural generation model, named HiStGen, for the task. Our model attempts to capture the three-level coherence via the following ways. First, we introduce a Markov paragraph dependency mechanism between context paragraphs for section identification. Second, we employ a section-aware attention mechanism to ensure the semantic coherence between a section and its heading. Finally, we leverage a Markov heading dependency mechanism and a review mechanism between context headings to improve the consistency and eliminate duplication between section headings. Besides, we build a novel WIKIOG dataset, a public collection which consists of over 1.75 million document-outline pairs for research on the OG task. Experimental results on our benchmark dataset demonstrate that our model can significantly outperform several state-of-the-art sequential generation models for the OG task.
http://arxiv.org/abs/1905.10039
Human flexible cognition and behavior indicate that the human brain can effectively use its internal feature representations acquired through limited experiences for new experiences in different domains. This function is analogous to transfer learning (TL) in the field of machine learning. TL uses a well-trained feature space in a specific task domain to improve performance in new tasks with insufficient training data. TL with rich feature representations, such as features of convolutional neural networks (CNNs), shows high generalization ability across different task domains. However, such TL is still insufficient in making machine learning attain generalization ability comparable to that of the human brain. To address this, we introduce a method for TL mediated by human brains to improve the performance of TL especially on pattern recognition in which human high-level cognition is considered. Our method transforms feature representations of audiovisual inputs in CNNs into those in activation patterns of individual brains via their association learned ahead using measured brain responses. Then, to estimate labels reflecting human cognition and behavior induced by the audiovisual inputs, the transformed representations are used for TL. We demonstrate that our brain-mediated TL (BTL) shows higher performance in the label estimation than the standard TL. In addition, we illustrate that the estimations mediated by different brains vary from brain to brain, and the variability reflects the individual variability in perception. Thus, our BTL provides a framework to improve the generalization ability of machine-learning feature representations and enable machine learning to estimate human-like cognition and behavior, including individual variability.
http://arxiv.org/abs/1905.10037
Existing personalized dialogue models use human designed persona descriptions to improve dialogue consistency. Collecting such descriptions from existing dialogues is expensive and requires hand-crafted feature designs. In this paper, we propose to extend Model-Agnostic Meta-Learning (MAML)(Finn et al., 2017) to personalized dialogue learning without using any persona descriptions. Our model learns to quickly adapt to new personas by leveraging only a few dialogue samples collected from the same user, which is fundamentally different from conditioning the response on the persona descriptions. Empirical results on Persona-chat dataset (Zhang et al., 2018) indicate that our solution outperforms non-meta-learning baselines using automatic evaluation metrics, and in terms of human-evaluated fluency and consistency.
http://arxiv.org/abs/1905.10033
Graph convolutional networks (GCNs) are powerful tools for graph-structured data. However, they have been recently shown to be prone to topological attacks. Despite substantial efforts to search for new architectures, it still remains a challenge to improve performance in both benign and adversarial situations simultaneously. In this paper, we re-examine the fundamental building block of GCN—the Laplacian operator—and highlight some basic flaws in the spatial and spectral domains. As an alternative, we propose an operator based on graph powering, and prove that it enjoys a desirable property of “spectral separation.” Based on the operator, we propose a robust learning paradigm, where the network is trained on a family of “‘smoothed” graphs that span a spatial and spectral range for generalizability. We also use the new operator in replacement of the classical Laplacian to construct an architecture with improved spectral robustness, expressivity and interpretability. The enhanced performance and robustness are demonstrated in extensive experiments.
https://arxiv.org/abs/1905.10029
European libraries and archives are filled with enciphered manuscripts from the early modern period. These include military and diplomatic correspondence, records of secret societies, private letters, and so on. Although they are enciphered with classical cryptographic algorithms, their contents are unavailable to working historians. We therefore attack the problem of automatically converting cipher manuscript images into plaintext. We develop unsupervised models for character segmentation, character-image clustering, and decipherment of cluster sequences. We experiment with both pipelined and joint models, and we give empirical results for multiple ciphers.
http://arxiv.org/abs/1810.04297
Temporal-difference learning (TD), coupled with neural networks, is among the most fundamental building blocks of deep reinforcement learning. However, due to the nonlinearity in value function approximation, such a coupling leads to nonconvexity and even divergence in optimization. As a result, the global convergence of neural TD remains unclear. In this paper, we prove for the first time that neural TD converges at a sublinear rate to the global optimum of the mean-squared projected Bellman error for policy evaluation. In particular, we show how such global convergence is enabled by the overparametrization of neural networks, which also plays a vital role in the empirical success of neural TD. Beyond policy evaluation, we establish the global convergence of neural (soft) Q-learning, which is further connected to that of policy gradient algorithms.
http://arxiv.org/abs/1905.10027
The quality of Neural Machine Translation (NMT) has been shown to significantly degrade when confronted with source-side noise. We present the first large-scale study of state-of-the-art English-to-German NMT on real grammatical noise, by evaluating on several Grammar Correction corpora. We present methods for evaluating NMT robustness without true references, and we use them for extensive analysis of the effects that different grammatical errors have on the NMT output. We also introduce a technique for visualizing the divergence distribution caused by a source-side error, which allows for additional insights.
http://arxiv.org/abs/1905.10024
Object detection has gained great progress driven by the development of deep learning. Compared with a widely studied task – classification, generally speaking, object detection even need one or two orders of magnitude more FLOPs (floating point operations) in processing the inference task. To enable a practical application, it is essential to explore effective runtime and accuracy trade-off scheme. Recently, a growing number of studies are intended for object detection on resource constraint devices, such as YOLOv1, YOLOv2, SSD, MobileNetv2-SSDLite, whose accuracy on COCO test-dev detection results are yield to mAP around 22-25% (mAP-20-tier). On the contrary, very few studies discuss the computation and accuracy trade-off scheme for mAP-30-tier detection networks. In this paper, we illustrate the insights of why RetinaNet gives effective computation and accuracy trade-off for object detection and how to build a light-weight RetinaNet. We propose to only reduce FLOPs in computational intensive layers and keep other layer the same. Compared with most common way – input image scaling for FLOPs-accuracy trade-off, the proposed solution shows a constantly better FLOPs-mAP trade-off line. Quantitatively, the proposed method result in 0.1% mAP improvement at 1.15x FLOPs reduction and 0.3% mAP improvement at 1.8x FLOPs reduction.
http://arxiv.org/abs/1905.10011
Early diagnosis of pulmonary nodules (PNs) can improve the survival rate of patients and yet is a challenging task for radiologists due to the image noise and artifacts in computed tomography (CT) images. In this paper, we propose a novel and effective abnormality detector implementing the attention mechanism and group convolution on 3D single-shot detector (SSD) called group-attention SSD (GA-SSD). We find that group convolution is effective in extracting rich context information between continuous slices, and attention network can learn the target features automatically. We collected a large-scale dataset that contained 4146 CT scans with annotations of varying types and sizes of PNs (even PNs smaller than 3mm were annotated). To the best of our knowledge, this dataset is the largest cohort with relatively complete annotations for PNs detection. Our experimental results show that the proposed group-attention SSD outperforms the classic SSD framework as well as the state-of-the-art 3DCNN, especially on some challenging lesion types.
http://arxiv.org/abs/1812.07166
This paper presents the first use of graph neural networks (GNNs) for higher-order proof search and demonstrates that GNNs can improve upon state-of-the-art results in this domain. Interactive, higher-order theorem provers allow for the formalization of most mathematical theories and have been shown to pose a significant challenge for deep learning. Higher-order logic is highly expressive and, even though it is well-structured with a clearly defined grammar and semantics, there still remains no well-established method to convert formulas into graph-based representations. In this paper, we consider several graphical representations of higher-order logic and evaluate them against the HOList benchmark for higher-order theorem proving.
http://arxiv.org/abs/1905.10006
Partially annotated clips contain rich temporal contexts that can complement the sparse key frame annotations in providing supervision for model training. We present a novel paradigm called Temporally-Adaptive Features (TAF) learning that can utilize such data to learn better single frame models. By imposing distinct temporal change rate constraints on different factors in the model, TAF enables learning from unlabeled frames using context to enhance model accuracy. TAF generalizes “slow feature” learning and we present much stronger empirical evidence than prior works, showing convincing gains for the challenging semantic segmentation task over a variety of architecture designs and on two popular datasets. TAF can be interpreted as an implicit label augmentation method but is a more principled formulation compared to existing explicit augmentation techniques. Our work thus connects two promising methods that utilize partially annotated clips for single frame model training and can inspire future explorations in this direction.
http://arxiv.org/abs/1905.10000
Visual Question Answering (VQA) deep-learning systems tend to capture superficial statistical correlations in the training data because of strong language priors and fail to generalize to test data with a significantly different question-answer (QA) distribution. To address this issue, we introduce a self-critical training objective that ensures that visual explanations of correct answers match the most influential image regions more than other competitive answer candidates. The influential regions are either determined from human visual/textual explanations or automatically from just significant words in the question and answer. We evaluate our approach on the VQA generalization task using the VQA-CP dataset, achieving a new state-of-the-art i.e. 49.5\% using textual explanations and 48.5\% using automatically annotated regions.
http://arxiv.org/abs/1905.09998
To distinguish the stem end and blossom end of navel orange from its black spot, we propose a real-time tiny detection model (RTTD) with low computational cost, compact architecture and high detection accuracy. In particular, based on the characteristics of the data, we apply pure dense connectivity to limit and simplify the design of the model architecture and use k-means clustering to set the size and aspect ratios of the default boxes. The architecture of model is based on deeply supervised object detectors (DSOD), and which reduces some components like dense block and prediction layers for efficient and adds some auxiliary structure like Squeeze-and-Excitation layer and Swish for accuracy. And we create a dataset in Pascal VOC format annotated the three types of detection targets stem end, blossom end and black spot. Experimental results on our orange data set confirm that RTTD has competitive results to the state-of-the-art one stage detectors like SSD, DSOD, YOLOv2, YOLOv3, RFB and FSSD, and it achieves 87.479%mAP at 131 FPS with only 5.812M parameters.
http://arxiv.org/abs/1905.09994
Swarm robotic search is concerned with searching targets in unknown environments (e.g., for search and rescue or hazard localization), using a large number of collaborating simple mobile robots. In such applications, decentralized swarm systems are touted for their task/coverage scalability, time efficiency, and fault tolerance. To guide the behavior of such swarm systems, two broad classes of approaches are available, namely nature-inspired swarm heuristics and multi-robotic search methods. However, simultaneously offering computationally-efficient scalability and fundamental insights into the exhibited behavior (instead of a black-box behavior model), remains challenging under either of these two class of approaches. In this paper, we develop an important extension of the batch Bayesian search method for application to embodied swarm systems, searching in a physical 2D space. Key contributions lie in: 1) designing an acquisition function that not only balances exploration and exploitation across the swarm, but also allows modeling knowledge extraction over trajectories; and 2) developing its distributed implementation to allow asynchronous task inference and path planning by the swarm robots. The resulting collective informative path planning approach is tested on target search case studies of varying complexity, where the target produces a spatially varying (measurable) signal. Significantly superior performance, in terms of mission completion efficiency, is observed compared to exhaustive search and random walk baselines, along with favorable performance scalability with increasing swarm size.
http://arxiv.org/abs/1905.09988
End-to-end trained neural networks (NNs) are a compelling approach to autonomous vehicle control because of their ability to learn complex tasks without manual engineering of rule-based decisions. However, challenging road conditions, ambiguous navigation situations, and safety considerations require reliable uncertainty estimation for the eventual adoption of full-scale autonomous vehicles. Bayesian deep learning approaches provide a way to estimate uncertainty by approximating the posterior distribution of weights given a set of training data. Dropout training in deep NNs approximates Bayesian inference in a deep Gaussian process and can thus be used to estimate model uncertainty. In this paper, we propose a Bayesian NN for end-to-end control that estimates uncertainty by exploiting feature map correlation during training. This approach achieves improved model fits, as well as tighter uncertainty estimates, than traditional element-wise dropout. We evaluate our algorithms on a challenging dataset collected over many different road types, times of day, and weather conditions, and demonstrate how uncertainties can be used in conjunction with a human controller in a parallel autonomous setting.
http://arxiv.org/abs/1805.04829
Next-generation ground-based solar observations require good image quality metrics for post-facto processing techniques. Based on the assumption that texture features in solar images are multi-fractal which can be extracted by a trained deep neural network as feature maps, a new reduced-reference objective image quality metric, the perception evaluation is proposed. The perception evaluation is defined as cosine distance of Gram matrix between feature maps extracted from high resolution reference image and that from blurred images. We evaluate performance of the perception evaluation with simulated and real observation images. The results show that with a high resolution image as reference, the perception evaluation can give robust estimate of image quality for solar images of different scenes.
http://arxiv.org/abs/1905.09980
Ensembling is a universally useful approach to boost the performance of machine learning models. However, individual models in an ensemble are typically trained independently in separate stages, without information access about the overall ensemble. In this paper, model ensembles are treated as first-class citizens, and their performance is optimized end-to-end with parameter sharing and a novel loss structure that improves generalization. On large-scale datasets including ImageNet, Youtube-8M, and Kinetics, we demonstrate a procedure that starts from a strongly performing single deep neural network, and constructs an EnsembleNet that has both a smaller size and better performance. Moreover, an EnsembleNet can be trained in one stage just like a single model without manual intervention.
http://arxiv.org/abs/1905.09979
Image retrieval is an important problem in the area of multimedia processing. This paper presents two new curvelet-based algorithms for texture retrieval which are suitable for use in constrained-memory devices. The developed algorithms are tested on three publicly available texture datasets: CUReT, Mondial-Marmi, and STex-fabric. Our experiments confirm the effectiveness of the proposed system. Furthermore, a weighted version of the proposed retrieval algorithm is proposed, which is shown to achieve promising results in the classification of seismic activities.
http://arxiv.org/abs/1905.09976
We propose a new framework for prototypical learning that bases decision-making on few relevant examples that we call prototypes. Our framework, which can be integrated into a wide range of neural network architectures including pre-trained models, utilizes an attention mechanism that relates the encoded representations to determine the prototypes. This results in a model that: (1)enables high-quality interpretability by outputting samples most relevant to the decision-making in addition to outputting the classification results; (2)allows state-of-the-art confidence estimation by quantifying the mismatch across prototype labels; (3)permits state-of-the-art detection of distribution mismatch; and (4)improves robustness to label noise. We demonstrate that our model can enable all these benefits without sacrificing accuracy.
http://arxiv.org/abs/1902.06292
We propose Shift R-CNN, a hybrid model for monocular 3D object detection, which combines deep learning with the power of geometry. We adapt a Faster R-CNN network for regressing initial 2D and 3D object properties and combine it with a least squares solution for the inverse 2D to 3D geometric mapping problem, using the camera projection matrix. The closed-form solution of the mathematical system, along with the initial output of the adapted Faster R-CNN are then passed through a final ShiftNet network that refines the result using our newly proposed Volume Displacement Loss. Our novel, geometrically constrained deep learning approach to monocular 3D object detection obtains top results on KITTI 3D Object Detection Benchmark, being the best among all monocular methods that do not use any pre-trained network for depth estimation.
http://arxiv.org/abs/1905.09970
Visual Tracking is a complex problem due to unconstrained appearance variations and dynamic environment. Extraction of complementary information from the object environment via multiple features and adaption to the target’s appearance variations are the key problems of this work. To this end, we propose a robust object tracking framework based on Unified Graph Fusion (UGF) of multi-cue to adapt to the object’s appearance. The proposed cross-diffusion of sparse and dense features not only suppresses the individual feature deficiencies but also extracts the complementary information from multi-cue. This iterative process builds robust unified features which are invariant to object deformations, fast motion, and occlusion. Robustness of the unified feature also enables the random forest classifier to precisely distinguish the foreground from the background, adding resilience to background clutter. In addition, we present a novel kernel-based adaptation strategy using outlier detection and a transductive reliability metric.
http://arxiv.org/abs/1812.06407
An emerging problem in trustworthy machine learning is to train models that produce robust interpretations for their predictions. We take a step towards solving this problem through the lens of axiomatic attribution of neural networks. Our theory is grounded in the recent work, Integrated Gradients (IG), in axiomatically attributing a neural network’s output change to its input change. We propose training objectives in classic robust optimization models to achieve robust IG attributions. Our objectives give principled generalizations of previous objectives designed for robust predictions, and they naturally degenerate to classic soft-margin training for one-layer neural networks. We also generalize previous theory and prove that the objectives for different robust optimization models are closely related. Experiments demonstrate the effectiveness of our method, and also point to intriguing problems which hint at the need for better optimization techniques or better neural network architectures for robust attribution training.
http://arxiv.org/abs/1905.09957
We address multi-modal trajectory forecasting of agents in unknown scenes by formulating it as a planning problem. We present an approach consisting of three models; a goal prediction model to identify potential goals of the agent, an inverse reinforcement learning model to plan optimal paths to each goal, and a trajectory generator to obtain future trajectories along the planned paths. Analysis of predictions on the Stanford drone dataset, shows generalizability of our approach to novel scenes.
http://arxiv.org/abs/1905.09949