ACML 2019 论文列表

Autonomous agents need to make decisions in a sequential manner, under partially observable environment, and in consideration of how other agents behave. In critical situations, such decisions need to be made in real time for example to avoid collisions and recover to safe conditions. We propose a technique of tree search where a deterministic and pessimistic scenario is used after a specified depth. Because there is no branching with the deterministic scenario, the proposed technique allows us to take into account the events that can occur far ahead in the future. The effectiveness of the proposed technique is demonstrated in Pommerman, a multi-agent environment used in a NeurIPS 2018 competition, where the agents that implement the proposed technique have won the first and third places.

End-to-end deep reinforcement learning has enabled agents to learn with little preprocessing by humans. However, it is still difficult to learn stably and efficiently because the learning method usually uses a nonlinear function approximation. Neural Episodic Control (NEC), which has been proposed in order to improve sample efficiency, is able to learn stably by estimating action values using a non-parametric method. In this paper, we propose an architecture that incorporates random projection into NEC to train with more stability. In addition, we verify the effectiveness of our architecture by Atari’s five games. The main idea is to reduce the number of parameters that have to learn by replacing neural networks with random projection in order to reduce dimensions while keeping the learning end-to-end.

Crowd flow forecasting plays an important role in risk assessment and public safety. It is a difficult task due to complex spatial-temporal dependencies as well as missing values in data. A number of models are proposed to predict crowd flow on city-scale, yet the missing pattern in city complex environment is seldomly considered. We propose a crowd flow forecasting model, Imputed Spatial-Temporal Convolution network(ISTC) to accurately predict the crowd flow in large complex buildings. ISTC uses convolution layers, whose structures are configured by graphs, to model the spatial-temporal correlations. Meanwhile ISTC adds imputation layers to handle the missing data. We demonstrate our model on several real data sets collected from sensors in a large six-floor commercial complex building. The results show that ISTC outperforms the baseline methods and is capable of handling data with as much as 40% missing data.

PAC-Bayesian set up involves a stochastic classifier characterized by a posterior distribution on a classifier set, offers a high probability bound on its averaged true risk and is robust to the training sample used. For a given posterior, this bound captures the trade off between averaged empirical risk and KL-divergence based model complexity term. Our goal is to identify an optimal posterior with the least PAC-Bayesian bound. We consider a finite classifier set and 5 distance functions: KL-divergence, its Pinsker’s and a sixth degree polynomial approximations; linear and squared distances. Linear distance based model results in a convex optimization problem and we obtain a closed form expression for its optimal posterior. For uniform prior, this posterior has full support with weights negative-exponentially proportional to number of misclassifications. Squared distance and Pinsker’s approximation bounds are possibly quasi-convex and are observed to have single local minimum. We derive fixed point equations (FPEs) using partial KKT system with strict positivity constraints. This obviates the combinatorial search for subset support of the optimal posterior. For uniform prior, exponential search on a full-dimensional simplex can be limited to an ordered subset of classifiers with increasing empirical risk values. These FPEs converge rapidly to a stationary point, even for a large classifier set when a solver fails. We apply these approaches to SVMs generated using a finite set of SVM regularization parameter values on 9 UCI datasets. The resulting optimal posteriors (on the set of regularization parameters) yield stochastic SVM classifiers with tight bounds. KL-divergence based bound is the tightest, but is computationally expensive due to its non-convex nature and multiple calls to a root finding algorithm. Optimal posteriors for all 5 distance functions have lowest 10% test error values on most datasets, with that of linear distance being the easiest to obtain.

Accurate access to nuclear information on digital pathology images can assist physicians in diagnosis and subsequent treatment. The pathological images have a large number of nuclei and part of nuclei is touching, manual segmentation is time consuming and error prone. Therefore it is an important task to develop a accurate nuclei segmentation method. For traditional methods, it is hard to obtain a accurately nuclei segmentation result, because the nuclei have many different characterizations. In this paper, we propose a new nuclei segmentation method (MDC-Net), which is a deep fully convolutional network. The network contains multiple residual operations to reduce detail loss in image. In addition, dilated convolution which has different dilation ratio is used to increase receptive field. MDC-Net contains the distance map and contour image, enhancing information on individual nuclei to get accurate segmentation results. We improve the segmentation effect by using the post-processing operate. We demonstrate that MDC-Net can obtain state-of-the-art results on public dataset with multiple organ slices compared with other popular methods.

Multivariate time series prediction has recently attracted extensive research attention due to its wide applications in the area of financial investment, energy consumption, environmental pollution and so on. Because of the temporal complexity and nonlinearity existing in multivariate time series, few existing models could provide satisfactory prediction results. In this paper, we proposed a novel prediction approach based on optimized temporal convolutional networks with stacked auto-encoders, which can achieve better prediction performance as demonstrated in the experiments. Stacked auto-encoders are employed to extract effective features from complex multivariate time series. A temporal convolutional network is then constructed serving as the prediction model, which has a flexible receptive field and enjoys faster training speed with parallel computing ability than recurrent neural networks. The optimal hyperparameters in these models are discovered by Bayesian optimization. We performed extensive experiments by comparing the proposed algorithms and other popular algorithms on three different datasets, where the proposed approach obtain the best prediction results in various prediction horizons. In addition, we carefully analyze the search process of Bayesian optimization and provide further insights into hyperparametric tuning processes combining the exploration strategy with the exploitation strategy.

Multiple Kernel Learning (MKL) algorithm effectively combines different kernels to improve the performance of classification. Most MKL algorithms implicitly map samples into feature space by the form of inner-product. In contrast, Multiple Empirical Kernel Learning (MEKL) can explicitly map the input spaces into feature spaces so that the mapped feature vectors are explicitly represented, which is easy to process and analyze the adaptability of kernels for input space. Meanwhile, in order to pay attention to the structure and discriminant information of samples in empirical feature space, inspired by discriminant locality preserving projections, we introduce the discriminant locality preservation regularization into MEKL framework to propose the Multiple Empirical Kernel Learning with Discriminant Locality Preservation (MEKL-DLP). Experiments conducted on real-world datasets validate the effectiveness of the proposed MEKL-DLP compared with the classical kernel-based algorithms and state-of-art MKL algorithms.

In this paper, we propose an end-to-end multi-width activation and multiple receptive field networks for the large-scale and complicated dynamic scene deblurring. Firstly, we design a multi-width activation feature extraction module, in which a multi-width activation residual block is proposed for making the activation function learn more the nonlinear information and extracting wider nonlinear features. Secondly, we design a multiple receptive field (RF) feature extraction module, in which a multiple RF residual block is proposed for enlarging the RF efficiently and capturing more nonlinear information from distant locations. And then, we design the multi-scale feature fusion module, where a learning fusion structure is designed to adaptively fuse the multi-scale features and complicated blur information from the different modules. Finally, we use a multi-component loss function to jointly optimize our networks. Extensive experimental results demonstrate that the proposed method outperforms the recent state-of-the-art deblurring methods, both quantitatively and qualitatively.

Successive POI recommendation is a fundamental problem for location-based social networks (LBSNs). POI recommendation takes a variety of POI context information (e.g. spatial location and textual comment) and user preference into consideration. Existing POI recommendation systems mainly focus on part of the POI context and user preference with a specific modeling, which loses valuable information from other aspects. In this paper, we propose to construct a multi-modal check-in graph, a heterogeneous graph that combines five check-in aspects in a unified way. We further propose a multi-modal representation learning model based on the graph to jointly learn POI and user representations. Finally, we employ an attentional recurrent neural network based on the representations for successive POI recommendation. Experiments on a public dataset studies the effects of modeling different aspects of check-in records and demonstrates the effectiveness of the method in improving POI recommendation performance.

Recently, Siamese networks have drawn great attention in the visual tracking community because of their balanced accuracy and speed. However, most existing Siamese frameworks describe the target appearance using a global pattern from the last layer, leading to high sensitivity to similar distractors, non-rigid appearance change, and partial occlusion. Addressing these issues, we propose a Multi-branch Siamese network (MSiam) for high-performance object tracking. The MSiam performs layer-wise feature aggregations and simultaneously considers the global-local patterns for more accurate target tracking. In particular, we propose a feature aggregation module (FAM) keeping the heterogeneity of the three types of features, further improving the discriminability of MSiam using both high-level semantic and low-level spatial information. To enhance the adaptability to non-rigid appearance change and partial occlusion, a multi-scale local pattern detection module (LPDM) is designed to identify discriminative regions of the target objects. By considering various combinations of the local structures, our tracker can form various types of structure patterns. Extensive evaluations on five benchmarks demonstrate that the proposed tracking algorithm performs favorably against state-of-the-art methods while running beyond real-time.

The bird community in the mangrove areas is an important component of the mangrove wetlands ecosystem and an indicator species for the assessment of the environmental health status of mangrove wetlands. The classification of bird species by the sound of bird in the mangrove areas has the advantages of less interference to the environment and wide monitoring range. In this paper, we propose a novel method that combines the feature recalibration mechanism with depthwise separable convolution for the mangrove bird sound classification. In the proposed method, we introduce Xception network in which depthwise separable convolution with lower parameter number and computational cost than traditional convolution can be stacked in a residual manner, as the baseline network. And we fuse the feature recalibration mechanism into the depthwise separable convolution for actively learning the weights of the feature channels in the network layer, so that we can enhance the important features in bird sound signals to improve the performance of the classification. In the proposed method, firstly we extract three-channel log-mel features of the bird sound signals and we introduce the mixup method to augment the extracted features. Secondly, we construct the recalibrated feature maps including the different scales of information to get the classification results. To verify the effectiveness of the proposed method, we build a dataset with 9282 samples including 25 kinds of the mangrove birds such as Egretta alba, Parus major, Charadrius dubius, etc. habiting in the mangroves of Fangcheng Port of China, and execute the experiments on the built dataset. Furthermore, we also validate the adaptability of our proposed method on the dataset of TAU Urban Acoustic Scenes 2019, and achieve a better result.

Multi-label learning learns from examples each associated with multiple class labels simultaneously, and the goal is to induce a predictive model which can assign a set of relevant labels for the unseen instance. Label-specific features serve as an effective strategy towards inducing multi-label predictive model, where the relevancy of each class label is determined by employing tailored features encoding inherent and distinct characteristics of the class label its own. In this paper, a regularization based approach named {\textsc{Reel}} is proposed for label-specific features generation, which works by enriching label-specific feature representation for each class label via synergizing informative label-specific features from other class labels with sparse regularization. Specifically, full-order label correlations are considered by {\textsc{Reel}} while the number of classifiers induced for multi-label prediction is linear to the number of class labels. Extensive experiments on fifteen benchmark multi-label data sets clearly show the favorable performance of {\textsc{Reel}} against other state-of-the-art multi-label learning approaches with label-specific features.

Leveraging an equivalence property in the state-space of a Markov Decision Process (MDP) has been investigated in several studies. This paper studies equivalence structure in the reinforcement learning (RL) setup, where transition distributions are no longer assumed to be known. We present a notion of similarity between transition probabilities of various state-action pairs of an MDP, which naturally defines an equivalence structure in the state-action space. We present equivalence-aware confidence sets for the case where the learner knows the underlying structure in advance. These sets are provably smaller than their corresponding equivalence-oblivious counterparts. In the more challenging case of an unknown equivalence structure, we present an algorithm called ApproxEquivalence that seeks to find an (approximate) equivalence structure, and define confidence sets using the approximate equivalence. To illustrate the efficacy of the presented confidence sets, we present C-UCRL, as a natural modification of UCRL2 for RL in undiscounted MDPs. In the case of a known equivalence structure, we show that C-UCRL improves over UCRL2 in terms of \emph{regret} by a factor of $\sqrt{SA/C}$, in any communicating MDP with $S$ states, $A$ actions, and $C$ classes, which corresponds to a massive improvement when $C\ll SA$. To the best of our knowledge, this is the first work providing regret bounds for RL when an equivalence structure in the MDP is efficiently exploited. In the case of an unknown equivalence structure, we show through numerical experiments that C-UCRL combined with ApproxEquivalence outperforms UCRL2 in ergodic MDPs.

We consider the online prediction of a varying Bernoulli process (sequence of varying Bernoulli probabilities) from a single binary sequence. A real-valued online prediction method has been proposed as a prior work that incorporates the smoothness of the prediction sequence into the concept of the regret. Also, a Bayesian prediction method for the varying Bernoulli processes has been developed based on the variational inference. However, the former is not applicable to loss functions other than the squared error function, and the latter has no guarantee on the regret as an online prediction method. We propose a new online prediction method of a varying Bernoulli process from a single binary sequence with a guarantee to minimize the maximum regret under variational approximation. Through numerical experiments, we compare the Bayesian prediction method with the proposed method by using the regret with/without approximation and the KL divergence from the true underlying process. We discuss the prediction accuracy and influences of the approximation of the proposed method.

\textit{Hard instances}, which require a long time for a specific algorithm to solve, help (1) analyze the algorithm for accelerating it and (2) build a good benchmark for evaluating the performance of algorithms. There exist several efforts for automatic generation of hard instances. For example, evolutionary algorithms have been utilized to generate hard instances. However, they generate only finite number of hard instances. The merit of such methods is limited because it is difficult to extract meaningful patterns from small number of instances. We seek for a probabilistic generator of hard instances. Once the generative distribution of hard instances is obtained, we can sample a variety of hard instances to build a benchmark, and we can extract meaningful patterns of hard instances from sampled instances. The existing methods for modeling the hard instance distribution rely on parameters or rules that are found by domain experts; however, they are specific to the problem. Hence, it is challenging to model the distribution for general cases. In this paper, we focus on graph problems. We propose \textsc{HiSampler}, the hard instance sampler, to model the hard instance distribution of graph algorithms. \textsc{HiSampler} makes it possible to obtain the distribution of hard instances without hand-engineered features. To the best of our knowledge, this is the first method to learn the distribution of hard instances using machine learning. Through experiments, we demonstrate that our proposed method can generate instances that are a few to several orders of magnitude harder than the random-based approach in many settings. In particular, our method outperforms rule-based algorithms in the 3-coloring problem.

Neural networks typically need huge amounts of data to train in order to get reasonable generalizable results. A common approach is to artificially generate samples by using prior knowledge of the data properties or other relevant domain knowledge. However, if the assumptions on the data properties are not accurate or the domain knowledge is irrelevant to the task at hand, one may end up degenerating learning performance by using such augmented data in comparison to simply training on the limited available dataset. We propose a critical data augmentation method using feature side-information, which is obtained from domain knowledge and provides detailed information about features' intrinsic properties. Most importantly, we introduce an instance wise quality checking procedure on the augmented data. It filters out irrelevant or harmful augmented data prior to entering the model. We validated this approach on both synthetic and real-world datasets, specifically in a scenario where the data augmentation is done based on a task independent, unreliable source of information. The experiments show that the introduced critical data augmentation scheme helps avoid performance degeneration resulting from incorporating wrong augmented data.

The problem of “learning to aggregate” (LTA) has recently been introduced as a novel machine learning setting, in which instances are represented in the form of a composition of a (variable) number on constituents. Such compositions are associated with an evaluation, which is the target of the prediction task, and which can presumably be modeled in the form of a suitable aggregation of the properties of its constituents. An especially interesting class of LTA problems arises when the evaluations of the constituents are not available at training time, and instead ought to be learned simultaneously with the aggregation function. This scenario is referred to as the “aggregation/disaggregation problem”. In this paper, we tackle this problem for an interesting type of aggregation function, namely the Ordered Weighted Averaging (OWA) operator. In particular, we provide an algorithm for learning the OWA parameters together with local utility scores of the constituents, and evaluate this algorithm in a case study on predicting the performance of classifier ensembles.

Nowadays, the top-$k$ accuracy is a major performance criterion when benchmarking multi-class classifier using datasets with a large number of categories. Top-$k$ multiclass SVM has been designed with the aim to minimize the empirical risk based on the top-$k$ accuracy. There already exist two SDCA-based algorithms to learn the top-$k$ SVM, enjoying several preferable properties for optimization, although both the algorithms suffer from two disadvantages. A weak point is that, since the design of the algorithms are specialized only to the top-$k$ hinge, their applicability to other variants is limited. The other disadvantage is that both the two algorithms cannot attain the optimal solution in most cases due to their theoritical imperfections. In this study, a weighted extension of top-$k$ SVM is considered, and novel learning algorithms based on the Frank-Wolfe algorithm is devised. The new learning algorithms possess all the favorable properties of SDCA as well as the applicability not only to the original top-$k$ SVM but also to the weighted extension. Geometrical convergence is achieved by smoothing the loss functions. Numerical simulations demonstrate that only the proposed Frank-Wolfe algorithms can converge to the optimum, in contrast with the failure of the two existing SDCA-based algorithms. Finally, our analytical results for these two studies are presented to shed light on the meaning of the solutions produced from their algorithms.

To explore underlying complementary information from multiple views, in this paper, we propose a novel Latent Multi-view Semi-Supervised Classification (LMSSC) method. Unlike most existing multi-view semi-supervised classification methods that learn the graph using original features, our method seeks an underlying latent representation and performs graph learning and label propagation based on the learned latent representation. With the complementarity of multiple views, the latent representation could depict the data more comprehensively than every single view individually, accordingly making the graph more accurate and robust as well. Finally, LMSSC integrates latent representation learning, graph construction, and label propagation into a unified framework, which makes each subtask optimized. Experimental results on real-world benchmark datasets validate the effectiveness of our proposed method.