ACML 2019 Papers

In many supervised learning settings, elicited labels comprise pairwise comparisons or rankings of samples. We propose a Bayesian inference model for ranking datasets, allowing us to take a probabilistic approach to ranking inference. Our probabilistic assumptions are motivated by, and consistent with, the so-called Plackett-Luce model. We propose a variational inference method to extract a closed-form Gaussian posterior distribution. We show experimentally that the resulting posterior yields more reliable ranking predictions compared to predictions via point estimates.

In this paper, we propose a novel variational generator framework for conditional GANs to catch semantic details for improving the generation quality and diversity. Traditional generators in conditional GANs simply concatenate the conditional vector with the noise as the input representation, which is directly employed for upsampling operations. However, the hidden condition information is not fully exploited, especially when the input is a class label. Therefore, we introduce a variational inference into the generator to infer the posterior of latent variable only from the conditional input, which helps achieve a variable augmented representation for image generation. Qualitative and quantitative experimental results show that the proposed method outperforms the state-of-the-art approaches and achieves the realistic controllable images.

Medical images from different modalities (e.g. MRI, CT) or contrasts (e.g. T1, T2) are usually used to extract abundant information for medical image analysis. Some modalities or contrasts may be degraded or missing, caused by artifacts or strict timing during acquisition. Thus synthesizing realistic medical images in the required domain is meaningful and helpful for clinical application. Meanwhile, due to the time-consuming of manual annotation, automatic medical image segmentation has attracted much attention. In this paper, we propose an end-to-end cross-domain synthesis and segmentation framework SSA-Net. It is based on cycle generative adversarial network (CycleGAN) for unpaired data. We introduce a gradient consistent term to refine the boundaries in synthesized images. Besides, we design a special shape consistent term to constrain the anatomical structure in synthesized images and to guide segmentation without target domian labels. In order to make the synthesis subnet focusing on some hard-to-learn regions automatically, we also introduce the attention block into the generator. On two challenging validation datasets (CHAOS and iSeg-2017), the proposed method achieves superior synthesis performance and comparable segmentation performance.

Recent years have witnessed significant progress in solving challenging problems across various domains using deep reinforcement learning (RL). Despite the success, the weak robustness has risen as a big obstacle for applying existing RL algorithms into real problems. In this paper, we propose unified policy optimization (UPO), a sample-efficient shared policy framework that allows a policy to update itself by considering different gradients generated by different policy gradient (PG) methods. Specifically, we propose two algorithms called UPO-MAB and UPO-ES, to leverage these different gradients by adopting the idea of multi-arm bandit (MAB) and evolution strategies (ES), with the purpose of finding the gradient direction leading to more performance gain with less extra data cost. Extensive experiments show that our approach can lead to stronger robustness and better performance than baselines.

Stochastic variational inference has emerged as an effective method for performing inference on or learning complex models for data. Yet, one of the challenges in stochastic variational inference is handling high-dimensional data, such as sequential data, and models with non-differentiable densities caused by, for instance, the use of discrete latent variables. In such cases, it is challenging to control the variance of the gradient estimator used in stochastic variational inference, while low variance is often one of the key properties needed for successful inference. In this work, we present a new algorithm for stochastic variational inference of sequential models which trades off bias for variance to tackle this challenge effectively. Our algorithm is inspired by variance reduction techniques in reinforcement learning, yet it uniquely adopts their key ideas in the context of stochastic variational inference. We demonstrate the effectiveness of our approach through formal analysis and experiments on synthetic and real-world datasets.

We tackle the blackbox issue of deep neural networks in the settings of reinforcement learning (RL) where neural agents learn towards maximizing reward gains in an uncontrollable way. Such learning approach is risky when the interacting environment includes an expanse of state space because it is then almost impossible to foresee all unwanted outcomes and penalize them with negative rewards beforehand. We propose Action-conditional $\beta$-VAE (AC-$\beta$-VAE) that allows succinct mappings of action-dependent factors in desirable dimensions of latent representations while disentangling environmental factors. Our proposed method tackles the blackbox issue by encouraging an RL policy network to learn interpretable latent features by distinguits influenshing ices from uncontrollable environmental factors, which closely resembles the way humans understand their scenes. Our experimental results show that the learned latent factors not only are interpretable, but also enable modeling the distribution of entire visited state-action space. We have experimented that this characteristic of the proposed structure can lead to ex post facto governance for desired behaviors of RL agents.

Can a text classifier generalize well for datasets where the text length is different? For example, when short reviews are sentiment-labeled, can these transfer to predict the sentiment of long reviews (i.e., short to long transfer), or vice versa? While unsupervised transfer learning has been well-studied for cross domain/lingual transfer tasks, \textbf{Cross Length Transfer} (CLT) has not yet been explored. One reason is the assumption that length difference is trivially transferable in classification. We show that it is not, because short/long texts differ in context richness and word intensity. We devise new benchmark datasets from diverse domains and languages, and show that existing models from similar tasks cannot deal with the unique challenge of transferring across text lengths. We introduce a strong baseline model called \textsc{BaggedCNN} that treats long texts as bags containing short texts. We propose a state-of-the-art CLT model called \textbf{Le}ngth \textbf{Tra}nsfer \textbf{Net}work\textbf{s} (\textsc{LeTraNets}) that introduces a two-way encoding scheme for short and long texts using multiple training mechanisms. We test our models and find that existing models perform worse than the \textsc{BaggedCNN} baseline, while \textsc{LeTraNets} outperforms all models.

A model segment method called Octree Subdivision has been presented for long years, which allows any three-dimensional point cloud object to be subdivided into infinitesimals so that it can be approximated by a particular surface function. In this paper, we proposed a new method named self-merging octree to reconstruct the surface of 3D Point Cloud which can be obtained by laser scanners or generated by some 3D modeling software. Different from any other surface reconstruction algorithms such as local property-based or specific type-based, a function pool-based was introduced in our research because it can express many different types of surfaces. We subdivide point cloud model by self-merging octree and categorize it by the neuro-network. In this idea, it is easy for us to find a proper surface function to present the subsurface of the model. What‘s more, while we extend the function pool, we can indicate far more style models. We have tried to reconstruct many point cloud models‘ surfaces in this way, and it works well and also shows its potential ability to build a bridge in the fields of model editing, model splicing, and model deformation.

We present an algorithm for learning decision trees using stochastic gradient information as the source of supervision. In contrast to previous approaches to gradient-based tree learning, our method operates in the incremental learning setting rather than the batch learning setting, and does not make use of soft splits or require the construction of a new tree for every update. We demonstrate how one can apply these decision trees to different problems by changing only the loss function, using classification, regression, and multi-instance learning as example applications. In the experimental evaluation, our method performs similarly to standard incremental classification trees, outperforms state of the art incremental regression trees, and achieves comparable performance with batch multi-instance learning methods.

In a software life cycle, bugs could happen at any time. Assigning bugs to relevant components/developers is a crucial task for software development. It is also a tough and resource consuming job. First, there are many components in a complex system and it is hard to understand their interactions and identify the root cause. Second, the list of components keeps growing for actively developed products and it is not easy to catch all updates. This task also faces several challenges from the machine learning point of view: 1) the ground truth is mixed with multiple levels of labels; 2) the data are severely imbalanced. 3). concept drift as future bugs are unlikely to come from the same distribution as the historical data. In this paper, we present a machine learning based solution for the bug assignment problem. We build component classifiers using a multi-layer Neural Network, based on features that were learned from data directly. A hierarchical classification framework is proposed to address the mixed label problem and improve the prediction accuracy. We also introduce a recency based sampling procedure to alleviate the data imbalance and concept drift problem. Our solution can easily accommodate new data and handle continuous system development/update.

In the past few years, facial expression recognition has made great progress because of the development of convolutional neural networks. However, the features learned only using the softmax loss are not discriminative enough for highly accurate facial expression recognition in the wild, especially for the compound facial expression recognition. To enhance the discriminative power of the learned features, we propose the separate loss for both basic and compound facial expression recognition in the wild in this paper. Such loss maximizes intra-class similarity while minimizing the similarity between different classes. The qualitative and quantitative analysis shows that the features learned using such loss function are characterized by intra-class compactness and inter-class separation. Experiments are performed on two databases in the wild and the proposed method achieves state-of-the-art results on both basic and compound expressions. Furthermore, another two databases are used to perform cross database experiments to show the generalization ability of our method.

Due to the efficiency of exploring multiple views of the real-word data, Multi-View Clustering (MVC) has attracted extensive attention from the scholars and researches based on it have made significant progress. However, multi-view data with numerous complementary information is vulnerable to various factors (such as noise). So it is an important and challenging task to discover the intrinsic characteristics hidden deeply in the data. In this paper, we present a novel MVC algorithm based on deep matrix factorization, named Self-Weighted Multi-view Clustering with Deep Matrix Factorization (SMDMF). By performing the deep decomposition structure, SMDMF can eliminate interference and reveal semantic information of the multi-view data. To properly integrate the complementary information among views, it assigns an automatic weight for each view without introducing supernumerary parameters. We also analyze the convergence of the algorithm and discuss the hierarchical parameters. The experimental results on four datasets show our algorithm is superior to other comparisons in all aspects.

As a new occurring unsupervised method, multi-view clustering offers a good way to investigate the hidden structure from multi-view data and attracts extensive attention in the community of machine learning and data mining. One popular approach is to identify a common latent subspace for capturing the multi-view information. However, these methods are still limited due to the unsupervised learning process and suffer from considerable noisy information from different views. To address this issue, we present a novel multi-view subspace clustering method, named self-supervised deep multi-view subspace clustering (\textbf{S2DMVSC}). It seamlessly integrates spectral clustering and affinity learning into a deep learning framework. \textbf{S2DMVSC} has two main merits. One is that the clustering results can be sufficiently exploited to supervise the latent representation learning for each view (via a classification loss) and the common latent subspace learning (via a spectral clustering loss) for multiple views. The other is that the affinity matrix among data objects is automatically computed according to the high-level and cluster-driven representation. Experiments on two scenarios, including original features and multiple hand-crafted features, demonstrate the superiority of the proposed approach over the state-of-the-art baselines.

The major challenge of learning from multi-label data has arisen from the overwhelming size of label space which makes this problem NP-hard. This problem can be alleviated by gradually involving easy to hard tags into the learning process. Besides, the utilization of a diversity maintenance approach avoids overfitting on a subset of easy labels. In this paper, we propose a self-paced multi-label learning with diversity (SPMLD) which aims to cover diverse labels with respect to its learning pace. In addition, the proposed framework is applied to an efficient correlation-based multi-label method. The non-convex objective function is optimized by an extension of the block coordinate descent algorithm. Empirical evaluations on real-world datasets with different dimensions of features and labels imply the effectiveness of the proposed predictive model.

Recovering high quality images from microscopic observations is an essential technology in biological imaging. Existing recovery methods require solving an optimization problem by using iterative algorithms, which are computationally expensive and time consuming. The focus of this study is to accelerate the image recovery by using deep neural networks (DNNs). In our approach, we first train a certain type of DNN by using some observations from microscopes, so that it can well approximate the image recovery process. The recovery of a new observation is then computed thorough a single forward propagation in the trained DNN. In this study, we specifically focus on observations obtained by SPoD (Super-resolution by Polarization Demodulation), a recently developed microscopic technique, and accelerate the image recovery for SPoD by using DNNs. To this end, we propose \emph{SPoD-Net}, a specifically tailored DNN for fast recovery of SPoD images. Unlike general DNNs, SPoD-Net can be parameterized using a small number of parameters, which is helpful in two ways: (i) it can be stored in a small memory, and (ii) it can be trained efficiently. We also propose a method to stabilize the training of SPoD-Net. In the experiments with the real SPoD observations, we confirmed the effectiveness of SPoD-Net over existing recovery methods. Specifically, we observed that SPoD-Net could recover images with more than a hundred times faster than the existing method.

Feature pyramid and feature fusing are widely used in object detection. Using feature pyramid can confront the challenge of scale variation across different objects. Feature fusing imports context information to improve detection performance. Although detecting with feature pyramid and feature fusing has achieved some encouraging results, there are still some limitations owing to the features’ level variance among different layers. In this paper, we exploit that serial-parallel combined feature fusing can enhance object detection. Instead of detecting on the feature pyramid of backbone directly, we fuse different layers from backbone as base features. Then the base features are fed into a U-shape module to build local-global feature pyramid. At last, we use the pyramid to do the multi-scale detection with our combined feature fusing method. We call this one-stage detector SPCDet. It keeps real time speed and outperforms other detectors in trade-off between accuracy and speed.

Nowadays the system vulnerability caused by soft errors grows exponentially, of which Silent Data Corruption(SDC) is one of the most harmful issues due to introducing unnoticed changes to the original data and error outputs. Thus, the detection of SDC-causing errors is extremely significant to the system reliability. However, most of the current detecting techniques require sufficient data of fault injections for training, which are difficult to achieve in practice because of high resources consumption, such as expensive execution time and code size costs. To this end, we propose a lightweight model named Deep Forest Regression based Multi-granularity Redundancy(DFRMR) to improve the error detection rate and meanwhile decrease the resources consumption. Specifically, first, we employ the program analysis to extract instruction features which are highly related to SDCs. Second, we design the deep forest regression model to predict the SDC vulnerability of instructions. Third, we optimize the error detection procedure by duplicating the critical instructions with different granularity. Finally, we evaluate our DFRMR model on Mibench benchmarks with multiple testing programs. The results show that our method attains better detection accuracy compared to other state-of-the-art methods and keeps the low multi-granularity redundancy.

The skip-connection and the batch-normalization (BN) in ResNet enable an extreme deep neural network to be trained with high performance. However, the reasons for its high performance are still unclear. To clear that, we study the effects of the skip-connection and the BN on the class-related signal propagation through hidden layers because a large ratio of the between-class distance to the within-class distance of feature vectors at the last hidden layer induces high performance. Our result shows that the between-class distance and the within-class distance change differently through layers: the deep multilayer perceptron with randomly initialized weights degrades the ratio of the between-class distance to the within-class distance and the skip-connection and the BN relax this degradation. Moreover, our analysis implies that the skip-connection and the BN encourage training to improve this distance ratio. These results imply that the skip-connection and the BN induce high performance.

We introduce a novel stochastic regularization technique for deep neural networks, which decomposes a layer into multiple branches with different parameters and merges stochastically sampled combinations of the outputs from the branches during training. Since the factorized branches can collapse into a single branch through a linear operation, inference requires no additional complexity compared to the ordinary layers. The proposed regularization method, referred to as StochasticBranch, is applicable to any linear layers such as fully-connected or convolution layers. The proposed regularizer allows the model to explore diverse regions of the model parameter space via multiple combinations of branches to find better local minima. An extensive set of experiments shows that our method effectively regularizes networks and further improves the generalization performance when used together with other existing regularization techniques.