ICML 2015 论文列表

We consider the problem of building probabilistic models that jointly model short natural language utterances and source code snippets. The aim is to bring together recent work on statistical modelling of source code and work on bimodal models of images and natural language. The resulting models are useful for a variety of tasks that involve natural language and source code. We demonstrate their performance on two retrieval tasks: retrieving source code snippets given a natural language query, and retrieving natural language descriptions given a source code query (i.e., source code captioning). The experiments show there to be promise in this direction, and that modelling the structure of source code is helpful towards the retrieval tasks.

We introduce BilBOWA (Bilingual Bag-of-Words without Alignments), a simple and computationally-efficient model for learning bilingual distributed representations of words which can scale to large monolingual datasets and does not require word-aligned parallel training data. Instead it trains directly on monolingual data and extracts a bilingual signal from a smaller set of raw-text sentence-aligned data. This is achieved using a novel sampled bag-of-words cross-lingual objective, which is used to regularize two noise-contrastive language models for efficient cross-lingual feature learning. We show that bilingual embeddings learned using the proposed model outperforms state-of-the-art methods on a cross-lingual document classification task as well as a lexical translation task on the WMT11 data.

We derive ensembles of decision trees through a nonparametric Bayesian model, allowing us to view such ensembles as samples from a posterior distribution. This insight motivates a class of Bayesian Forest (BF) algorithms that provide small gains in performance and large gains in interpretability. Based on the BF framework, we are able to show that high-level tree hierarchy is stable in large samples. This motivates an empirical Bayesian Forest (EBF) algorithm for building approximate BFs on massive distributed datasets and we show that EBFs outperform sub-sampling based alternatives by a large margin.

We consider the problem of quickly localizing multiple targets by asking questions of the form “How many targets are within this set" while obtaining noisy answers. This setting is a generalization to multiple targets of the game of 20 questions in which only a single target is queried. We assume that the targets are points on the real line, or in a two dimensional plane for the experiments, drawn independently from a known distribution. We evaluate the performance of a policy using the expected entropy of the posterior distribution after a fixed number of questions with noisy answers. We derive a lower bound for the value of this problem and study a specific policy, named the dyadic policy. We show that this policy achieves a value which is no more than twice this lower bound when answers are noise-free, and show a more general constant factor approximation guarantee for the noisy setting. We present an empirical evaluation of this policy on simulated data for the problem of detecting multiple instances of the same object in an image. Finally, we present experiments on localizing multiple faces simultaneously on real images.

Training Deep Neural Networks is complicated by the fact that the distribution of each layer’s inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs. Our method draws its strength from making normalization a part of the model architecture and performing the normalization for each training mini-batch. Batch Normalization allows us to use much higher learning rates and be less careful about initialization, and in some cases eliminates the need for Dropout. Applied to a stateof-the-art image classification model, Batch Normalization achieves the same accuracy with 14 times fewer training steps, and beats the original model by a significant margin. Using an ensemble of batch-normalized networks, we improve upon the best published result on ImageNet classification: reaching 4.82% top-5 test error, exceeding the accuracy of human raters.

We consider a budgeted learning setting, where the learner can only choose and observe a small subset of the attributes of each training example. We develop efficient algorithms for Ridge and Lasso linear regression, which utilize the geometry of the data by a novel distribution-dependent sampling scheme, and have excess risk bounds which are better a factor of up to O(d/k) over the state-of-the-art, where d is the dimension and k+1 is the number of observed attributes per example. Moreover, under reasonable assumptions, our algorithms are the first in our setting which can provably use *less* attributes than full-information algorithms, which is the main concern in budgeted learning. We complement our theoretical analysis with experiments which support our claims.

The emergence of compact GPS systems and the establishment of open data initiatives has resulted in widespread availability of spatial data for many urban centres. These data can be leveraged to develop data-driven intelligent resource allocation systems for urban issues such as policing, sanitation, and transportation. We employ techniques from Bayesian non-parametric statistics to develop a process which captures a common characteristic of urban spatial datasets. Specifically, our new spatial process framework models events which occur repeatedly at discrete spatial points, the number and locations of which are unknown a priori. We develop a representation of our spatial process which facilitates posterior simulation, resulting in an interpretable and computationally tractable model. The framework’s superiority over both empirical grid-based models and Dirichlet process mixture models is demonstrated by fitting, interpreting, and comparing models of graffiti prevalence for both downtown Vancouver and Manhattan.

We introduce a novel Gaussian process based Bayesian model for asymmetric transfer learning. We adopt a two-layer feed-forward deep Gaussian process as the task learner of source and target domains. The first layer projects the data onto a separate non-linear manifold for each task. We perform knowledge transfer by projecting the target data also onto the source domain and linearly combining its representations on the source and target domain manifolds. Our approach achieves the state-of-the-art in a benchmark real-world image categorization task, and improves on it in cross-tissue tumor detection from histopathology tissue slide images.

This paper provides an analysis of error propagation in Approximate Dynamic Programming applied to zero-sum two-player Stochastic Games. We provide a novel and unified error propagation analysis in L_p-norm of three well-known algorithms adapted to Stochastic Games (namely Approximate Value Iteration, Approximate Policy Iteration and Approximate Generalized Policy Iteration). We show that we can achieve a stationary policy which is \frac2γ(1 - γ)^2 ε+ \frac1(1 - γ)^2ε’-optimal, where εis the value function approximation error and ε’ is the approximate greedy operator error. In addition, we provide a practical algorithm (AGPI-Q) to solve infinite horizon γ-discounted two-player zero-sum stochastic games in a batch setting. It is an extension of the Fitted-Q algorithm (which solves Markov Decisions Processes in a batch setting) and can be non-parametric. Finally, we demonstrate experimentally the performance of AGPI-Q on a simultaneous two-player game, namely Alesia.

The growing need for labeled training data has made crowdsourcing an important part of machine learning. The quality of crowdsourced labels is, however, adversely affected by three factors: (1) the workers are not experts; (2) the incentives of the workers are not aligned with those of the requesters; and (3) the interface does not allow workers to convey their knowledge accurately, by forcing them to make a single choice among a set of options. In this paper, we address these issues by introducing approval voting to utilize the expertise of workers who have partial knowledge of the true answer, and coupling it with a ("strictly proper") incentive-compatible compensation mechanism. We show rigorous theoretical guarantees of optimality of our mechanism together with a simple axiomatic characterization. We also conduct preliminary empirical studies on Amazon Mechanical Turk which validate our approach.

Zero-shot learning consists in learning how to recognize new concepts by just having a description of them. Many sophisticated approaches have been proposed to address the challenges this problem comprises. In this paper we describe a zero-shot learning approach that can be implemented in just one line of code, yet it is able to outperform state of the art approaches on standard datasets. The approach is based on a more general framework which models the relationships between features, attributes, and classes as a two linear layers network, where the weights of the top layer are not learned but are given by the environment. We further provide a learning bound on the generalization error of this kind of approaches, by casting them as domain adaptation methods. In experiments carried out on three standard real datasets, we found that our approach is able to perform significantly better than the state of art on all of them, obtaining a ratio of improvement up to 17%.

We investigate the bilinear model, which is a matrix form linear model with the rank 1 constraint. A new online learning algorithm is proposed to train the model parameters. Our algorithm runs in the manner of online mirror descent, and gradients are computed by the power iteration. To analyze it, we give a new second order approximation of the squared spectral norm, which helps us to get a regret bound. Experiments on two sequential labelling tasks give positive results.

In this paper, we consider the problem of column subset selection. We present a novel analysis of the spectral norm reconstruction for a simple randomized algorithm and establish a new bound that depends explicitly on the sampling probabilities. The sampling dependent error bound (i) allows us to better understand the tradeoff in the reconstruction error due to sampling probabilities, (ii) exhibits more insights than existing error bounds that exploit specific probability distributions, and (iii) implies better sampling distributions. In particular, we show that a sampling distribution with probabilities proportional to the square root of the statistical leverage scores is better than uniform sampling, and is better than leverage-based sampling when the statistical leverage scores are very nonuniform. And by solving a constrained optimization problem related to the error bound with an efficient bisection search we are able to achieve better performance than using either the leverage-based distribution or that proportional to the square root of the statistical leverage scores. Numerical simulations demonstrate the benefits of the new sampling distributions for low-rank matrix approximation and least square approximation compared to state-of-the art algorithms.

Stochastic variational inference (SVI) is emerging as the most promising candidate for scaling inference in Bayesian probabilistic models to large datasets. However, the performance of these methods has been assessed primarily in the context of Bayesian topic models, particularly latent Dirichlet allocation (LDA). Deriving several new algorithms, and using synthetic, image and genomic datasets, we investigate whether the understanding gleaned from LDA applies in the setting of sparse latent factor models, specifically beta process factor analysis (BPFA). We demonstrate that the big picture is consistent: using Gibbs sampling within SVI to maintain certain posterior dependencies is extremely effective. However, we also show that different posterior dependencies are important in BPFA relative to LDA.

The Recurrent Neural Network (RNN) is an extremely powerful sequence model that is often difficult to train. The Long Short-Term Memory (LSTM) is a specific RNN architecture whose design makes it much easier to train. While wildly successful in practice, the LSTM’s architecture appears to be ad-hoc so it is not clear if it is optimal, and the significance of its individual components is unclear. In this work, we aim to determine whether the LSTM architecture is optimal or whether much better architectures exist. We conducted a thorough architecture search where we evaluated over ten thousand different RNN architectures, and identified an architecture that outperforms both the LSTM and the recently-introduced Gated Recurrent Unit (GRU) on some but not all tasks. We found that adding a bias of 1 to the LSTM’s forget gate closes the gap between the LSTM and the GRU.

We propose a distributed first-order augmented Lagrangian (DFAL) algorithm to minimize the sum of composite convex functions, where each term in the sum is a private cost function belonging to a node, and only nodes connected by an edge can directly communicate with each other. This optimization model abstracts a number of applications in distributed sensing and machine learning. We show that any limit point of DFAL iterates is optimal; and for any eps > 0, an eps-optimal and eps-feasible solution can be computed within O(log(1/eps)) DFAL iterations, which require O(\psi_\textmax^1.5/d_\textmin ⋅1/ε) proximal gradient computations and communications per node in total, where \psi_\textmax denotes the largest eigenvalue of the graph Laplacian, and d_\textmin is the minimum degree of the graph. We also propose an asynchronous version of DFAL by incorporating randomized block coordinate descent methods; and demonstrate the efficiency of DFAL on large scale sparse-group LASSO problems.

In this paper, we develop a new entropic matching kernel for weighted graphs by aligning depth-based representations. We demonstrate that this kernel can be seen as an \textbfaligned subtree kernel that incorporates explicit subtree correspondences, and thus addresses the drawback of neglecting the relative locations between substructures that arises in the R-convolution kernels. Experiments on standard datasets demonstrate that our kernel can easily outperform state-of-the-art graph kernels in terms of classification accuracy.

Although recent work in non-linear dimensionality reduction investigates multiple choices of divergence measure during optimization \citeyang2013icml,bunte2012neuro, little work discusses the direct effects that divergence measures have on visualization. We study this relationship, theoretically and through an empirical analysis over 10 datasets. Our works shows how the αand βparameters of the generalized alpha-beta divergence can be chosen to discover hidden macro-structures (categories, e.g. birds) or micro-structures (fine-grained classes, e.g. toucans). Our method, which generalizes t-SNE \citetsne, allows us to discover such structure without extensive grid searches over (α, β) due to our theoretical analysis: such structure is apparent with particular choices of (α, β) that generalize across datasets. We also discuss efficient parallel CPU and GPU schemes which are non-trivial due to the tree-structures employed in optimization and the large datasets that do not fully fit into GPU memory. Our method runs 20x faster than the fastest published code \citefmm. We conclude with detailed case studies on the following very large datasets: ILSVRC 2012, a standard computer vision dataset with 1.2M images; SUSY, a particle physics dataset with 5M instances; and HIGGS, another particle physics dataset with 11M instances. This represents the largest published visualization attained by SNE methods. We have open-sourced our visualization code: \texttthttp://rll.berkeley.edu/absne/.

We address the pre-image problem encountered in structured output prediction and the one of finding a string maximizing the prediction function of various kernel-based classifiers and regressors. We demonstrate that these problems reduce to a common combinatorial problem valid for many string kernels. For this problem, we propose an upper bound on the prediction function which has low computational complexity and which can be used in a branch and bound search algorithm to obtain optimal solutions. We also show that for many string kernels, the complexity of the problem increases significantly when the kernel is normalized. On the optical word recognition task, the exact solution of the pre-image problem is shown to significantly improve the prediction accuracy in comparison with an approximation found by the best known heuristic. On the task of finding a string maximizing the prediction function of kernel-based classifiers and regressors, we highlight that existing methods can be biased toward long strings that contain many repeated symbols. We demonstrate that this bias is removed when using normalized kernels. Finally, we present results for the discovery of lead compounds in drug discovery. The source code can be found at https://github.com/a-ro/preimage