ACML 2017 论文列表

Big Data problems in Machine Learning have large number of data points or large number of features, or both, which make training of models difficult because of high computational complexities of single iteration of learning algorithms. To solve such learning problems, Stochastic Approximation offers an optimization approach to make complexity of each iteration independent of number of data points by taking only one data point or mini-batch of data points during each iteration and thereby helping to solve problems with large number of data points. Similarly, Coordinate Descent offers another optimization approach to make iteration complexity independent of the number of features/coordinates/variables by taking only one feature or block of features, instead of all, during an iteration and thereby helping to solve problems with large number of features. In this paper, an optimization framework, namely, Batch Block Optimization Framework has been developed to solve big data problems using the best of Stochastic Approximation as well as the best of Coordinate Descent approaches, independent of any solver. This framework is used to solve strongly convex and smooth empirical risk minimization problem with gradient descent (as a solver) and two novel Stochastic Average Adjusted Gradient methods have been proposed to reduce variance in mini-batch and block-coordinate setting of the developed framework. Theoretical analysis prove linear convergence of the proposed methods and empirical results with bench marked datasets prove the superiority of proposed methods against existing methods.

In this paper, we present a novel method for solving structured prediction problems, based on combining Input Output Kernel Regression (IOKR) with an extension of magnitude-preserving ranking to structured output spaces. In particular, we concentrate on the case where a set of candidate outputs has been given, and the associated pre-image problem calls for ranking the set of candidate outputs. Our method, called magnitude-preserving IOKR, both aims to produce a good approximation of the output feature vectors, and to preserve the magnitude differences of the output features in the candidate sets. For the case where the candidate set does not contain corresponding ’correct’ inputs, we propose a method for approximating the inputs through application of IOKR in the reverse direction. We apply our method to two learning problems: cross-lingual document retrieval and metabolite identification. Experiments show that the proposed approach improves performance over IOKR, and in the latter application obtains the current state-of-the-art accuracy.

Convolutional Neural Networks (CNN) and the locally connected layer are limited in capturing the importance and relations of different local receptive fields, which are often crucial for tasks such as face verification, visual question answering, and word sequence prediction. To tackle the issue, we propose a novel locally smoothed neural network (LSNN) in this paper. The main idea is to represent the weight matrix of the locally connected layer as the product of the kernel and the smoother, where the kernel is shared over different local receptive fields, and the smoother is for determining the importance and relations of different local receptive fields. Specifically, a multi-variate Gaussian function is utilized to generate the smoother, for modeling the location relations among different local receptive fields. Furthermore, the content information can also be leveraged by setting the mean and precision of the Gaussian function according to the content. Experiments on some variant of MNIST clearly show our advantages over CNN and locally connected layer.

End-to-end learning refers to training a possibly complex learning system by applying gradient-based learning to the system as a whole. End-to-end learning systems are specifically designed so that all modules are differentiable. In effect, not only a central learning machine, but also all “peripheral” modules like representation learning and memory formation are covered by a holistic learning process. The power of end-to-end learning has been demonstrated on many tasks, like playing a whole array of Atari video games with a single architecture. While pushing for solutions to more challenging tasks, network architectures keep growing more and more complex. In this paper we ask the question whether and to what extent end-to-end learning is a future-proof technique in the sense of \emphscaling to complex and diverse data processing architectures. We point out potential inefficiencies, and we argue in particular that end-to-end learning does not make optimal use of the modular design of present neural networks. Our surprisingly simple experiments demonstrate these inefficiencies, up to the complete breakdown of learning.

By exploiting the property that the RBM log-likelihood function is the difference of convex functions, we formulate a stochastic variant of the difference of convex functions (DC) programming to minimize the negative log-likelihood. Interestingly, the traditional contrastive divergence algorithm is a special case of the above formulation and the hyperparameters of the two algorithms can be chosen such that the amount of computation per mini-batch is identical. We show that for a given computational budget the proposed algorithm almost always reaches a higher log-likelihood more rapidly, compared to the standard contrastive divergence algorithm. Further, we modify this algorithm to use the centered gradients and show that it is more efficient and effective compared to the standard centered gradient algorithm on benchmark datasets.

We present a new method for forecasting systems of multiple interrelated time series. The method learns the forecast models together with discovering leading indicators from within the system that serve as good predictors improving the forecast accuracy and a cluster structure of the predictive tasks around these. The method is based on the classical linear vector autoregressive model (VAR) and links the discovery of the leading indicators to inferring sparse graphs of Granger causality. We formulate a new constrained optimisation problem to promote the desired sparse structures across the models and the sharing of information amongst the learning tasks in a multi-task manner. We propose an algorithm for solving the problem and document on a battery of synthetic and real-data experiments the advantages of our new method over baseline VAR models as well as the state-of-the-art sparse VAR learning methods.

Deep learning methods have been actively researched for cross-modal retrieval, with the softmax cross-entropy loss commonly applied for supervised learning. However, the softmax cross-entropy loss is known to result in large intra-class variances, which is not not very suited for cross-modal matching. In this paper, a deep architecture called Deep Semantic Embedding (DSE) is proposed, which is trained in an end-to-end manner for image-text cross-modal retrieval. With images and texts mapped to a feature embedding space, class labels are used to guide the embedding learning, so that the embedding space has a semantic meaning common for both images and texts. This way, the difference between different modalities is eliminated. Under this framework, the center loss is introduced beyond the commonly used softmax cross-entropy loss to achieve both inter-class separation and intra-class compactness. Besides, a distance based softmax cross-entropy loss is proposed to jointly consider the softmax cross-entropy and center losses in fully gradient based learning. Experiments have been done on three popular image-text cross-modal retrieval databases, showing that the proposed algorithms have achieved the best overall performances.

Convolutional neural networks (CNNs) have yielded the excellent performance in a variety of computer vision tasks, where CNNs typically adopt a similar structure consisting of convolution layers, pooling layers and fully connected layers. In this paper, we propose to apply a novel method, namely Hybrid Orthogonal Projection and Estimation (HOPE), to CNNs in order to introduce orthogonality into the CNN structure. The HOPE model can be viewed as a hybrid model to combine feature extraction using orthogonal linear projection with mixture models. It is an effective model to extract useful information from the original high-dimension feature vectors and meanwhile filter out irrelevant noises. In this work, we present three different ways to apply the HOPE models to CNNs, i.e., \em HOPE-Input, \em single-HOPE-Block and \em multi-HOPE-Blocks. For \em HOPE-Input CNNs, a HOPE layer is directly used right after the input to de-correlate high-dimension input feature vectors. Alternatively, in \em single-HOPE-Block and \em multi-HOPE-Blocks CNNs, we consider to use HOPE layers to replace one or more blocks in the CNNs, where one block may include several convolutional layers and one pooling layer. The experimental results on CIFAR-10, CIFAR-100 and ImageNet databases have shown that the orthogonal constraints imposed by the HOPE layers can significantly improve the performance of CNNs in these image classification tasks (we have achieved one of the best performance when image augmentation has not been applied, and top 5 performance with image augmentation).

With the fast development of data collection techniques, a huge amount of complex multi-modal data are generated, shared and stored on the Internet. The burden of extracting multi-modal features for test instances in data analysis becomes the main fact that hurts the efficiency of prediction. In this paper, in order to reduce the modal extraction cost in serialized classification system, we propose a novel end-to-end serialized adaptive decision approach named Discriminative Modal Pursuit (\sc Dmp), which can automatically extract instance-specifically discriminative modal sequence for reducing the cost of feature extraction in the test phase. Rather than jointly optimize a highly non-convex empirical risk minimization problem, we are inspired by LSTM, and the proposed \sc Dmp can turn to learn the decision policies which predict the label information and decide the modalities to be extracted simultaneously within limited modal acquisition budget. Consequently, \sc Dmp approach can balance the classification performance and modal feature extraction cost by utilizing different modalities for different test instances. Empirical studies show that \sc Dmp is more efficient and effective than existing modal/feature extraction methods.

Regularized estimators in the context of group variables have been applied successfully in model and feature selection in order to preserve interpretability. We formulate a Distributionally Robust Optimization (DRO) problem which recovers popular estimators, such as Group Square Root Lasso (GSRL). Our DRO formulation allows us to interpret GSRL as a game, in which we learn a regression parameter while an adversary chooses a perturbation of the data. We wish to pick the parameter to minimize the expected loss under any plausible model chosen by the adversary - who, on the other hand, wishes to increase the expected loss. The regularization parameter turns out to be precisely determined by the amount of perturbation on the training data allowed by the adversary. In this paper, we introduce a data-driven (statistical) criterion for the optimal choice of regularization, which we evaluate asymptotically, in closed form, as the size of the training set increases. Our easy-to-evaluate regularization formula is compared against cross-validation, showing comparable performance.

In the design of deep neural architectures, recent studies have demonstrated the benefits of grouping subnetworks into a larger network. For examples, the Inception architecture integrates multi-scale subnetworks and the residual network can be regarded that a residual unit combines a residual subnetwork with an identity shortcut. In this work, we embrace this observation and propose the Competitive Pathway Network (CoPaNet). The CoPaNet comprises a stack of competitive pathway units and each unit contains multiple parallel residual-type subnetworks followed by a max operation for feature competition. This mechanism enhances the model capability by learning a variety of features in subnetworks. The proposed strategy explicitly shows that the features propagate through pathways in various routing patterns, which is referred to as pathway encoding of category information. Moreover, the cross-block shortcut can be added to the CoPaNet to encourage feature reuse. We evaluated the proposed CoPaNet on four object recognition benchmarks: CIFAR-10, CIFAR-100, SVHN, and ImageNet. CoPaNet obtained the state-of-the-art or comparable results using similar amounts of parameters. The code of CoPaNet is available at: \urlhttps://github.com/JiaRenChang/CoPaNet.

Leveraging the celebrated support vector regression (SVR) method, we propose a unifying framework in order to deliver regression machines in reproducing kernel Hilbert spaces (RKHSs) with data sparsity. The central point is a new definition of $ε$-insensitivity, valid for many regression losses (including quantile and expectile regression) and their multivariate extensions. We show that the dual optimization problem to empirical risk minimization with $ε$-insensitivity involves a data sparse regularization. We also provide an analysis of the excess of risk as well as a randomized coordinate descent algorithm for solving the dual. Numerical experiments validate our approach.

Sequence modeling with neural networks has lead to powerful models of symbolic music data. We address the problem of exploiting these models to reach creative musical goals, by combining with human input. To this end we generalise previous work, which sampled Markovian sequence models under the constraint that the sequence belong to the language of a given finite state machine provided by the human. We consider more expressive non-Markov models, thereby requiring approximate sampling which we provide in the form of an efficient sequential Monte Carlo method. In addition we provide and compare with a beam search strategy for conditional probability maximisation. Our algorithms are capable of convincingly re-harmonising famous musical works. To demonstrate this we provide visualisations, quantitative experiments, a human listening test and audio examples. We find both the sampling and optimisation procedures to be effective, yet complementary in character. For the case of highly permissive constraint sets, we find that sampling is to be preferred due to the overly regular nature of the optimisation based results. The generality of our algorithms permits countless other creative applications.

Knowledge graphs (KGs) are often significantly incomplete, necessitating a demand for KG completion. Path-based relation inference is one of the most important approaches to this task. Traditional methods treat each path between entity pairs as an atomic feature, thus inducing sparsity. Recently, neural network models solve this problem by decomposing a path as the sequence of relations in the path, before modelling path representations with Recurrent Neural Network (RNN) architectures. In cases there are multiple paths between an entity pair, state-of-the-art neural models either select only one path, or make usage of simple score pooling methods like Top-K, Average, LogSumExp. Unfortunately, none of these methods can model the scenario where relations can only be inferred by considering multiple informative paths collectively. In this paper, we propose a novel path-based relation inference model that learns entity pair representations with attentive path combination. Given an entity pair and a set of paths connecting the pair, our model allows for integrating information from each informative path, and form a dynamic entity pair representation for each query relation. We empirically evaluate the proposed method on a real-world dataset. Experimental results show that the proposed model achieves better performance than state-of-the-art path-based relation inference methods.

Imbalanced data poses a serious challenge for many machine learning and data mining applications. It may significantly affect the performance of learning algorithms. In digital marketing applications, events of interest (positive instances for building predictive models) such as click and purchase are rare. A retail website can easily receive a million visits every day, yet only a small percentage of visits lead to purchase. The large amount of raw data and the small percentage of positive instances make it challenging to build decent predictive models in a timely fashion. In this paper, we propose an adaptive sampling strategy to deal with this problem. It efficiently returns high quality training data, ensures system responsiveness and improves predictive performances.

This work addresses the instability in asynchronous data parallel optimization. It does so by introducing a novel distributed optimizer which is able to efficiently optimize a centralized model under communication constraints. The optimizer achieves this by pushing a normalized sequence of first-order gradients to a parameter server. This implies that the magnitude of a worker delta is smaller compared to an accumulated gradient, and provides a better direction towards a minimum compared to first-order gradients, which in turn also forces possible implicit momentum fluctuations to be more aligned since we make the assumption that all workers contribute towards a single minima. As a result, our approach mitigates the parameter staleness problem more effectively since staleness in asynchrony induces (implicit) momentum, and achieves a better convergence rate compared to other optimizers such as asynchronous \textsceasgd and \textscdynsgd, which we show empirically.

In natural language processing and related fields, it has been shown that the word embeddings can successfully capture both the semantic and syntactic features of words. They can serve as complementary information to topics models, especially for the cases where word co-occurrence data is insufficient, such as with short texts. In this paper, we propose a focused topic model where how a topic focuses on words is informed by word embeddings. Our models is able to discover more informed and focused topics with more representative words, leading to better modelling accuracy and topic quality. With the data argumentation technique, we can derive an efficient Gibbs sampling algorithm that benefits from the fully local conjugacy of the model. We conduct extensive experiments on several real world datasets, which demonstrate that our model achieves comparable or improved performance in terms of both perplexity and topic coherence, particularly in handling short text data.

The main task in training a linear classifier is to solve an unconstrained minimization problem. To apply an optimization method typically we iteratively find a good direction and then decide a suitable step size. Past developments of extending optimization methods for large-scale linear classification focus on finding the direction, but little attention has been paid on adjusting the step size. In this work, we explain that inappropriate step-size adjustment may lead to serious slow convergence. Among the two major methods for step-size selection, line search and trust region, we focus on investigating the trust region methods. After presenting some detailed analysis, we develop novel and effective techniques to adjust the trust-region size. Experiments indicate that our new settings significantly outperform existing implementations for large-scale linear classification.

We propose a Quantum-Inspired Subspace(QIS) Ensemble Method for generating feature ensembles based on feature selections. We assign each principal component a Fraction Transition Probability as its probability weight based on Principal Component Analysis and quantum interpretations. In order to generate the feature subset for each base regressor, we select a feature subset from principal components based on Fraction Transition Probabilities. The idea originating from quantum mechanics can encourage ensemble diversity and the accuracy simultaneously. We incorporate Quantum-Inspired Subspace Method into Random Forest and propose Quantum-Inspired Forest. We theoretically prove that the quantum interpretation corresponds to the first order approximation of ensemble regression. We also evaluate the empirical performance of Quantum-Inspired Forest and Random Forest in multiple hyperparameter settings. Quantum-Inspired Forest proves the significant robustness of the default hyperparameters on most data sets. The contribution of this work is two-fold, a novel ensemble regression algorithm inspired by quantum mechanics and the theoretical connection between quantum interpretations and machine learning algorithms.