ACML 2021 论文列表

Generative models with both discrete and continuous latent variables are highly motivated by the structure of many real-world data sets. They present, however, subtleties in training often manifesting in the discrete latent variable not being leveraged. In this paper, we show why such models struggle to train using traditional log-likelihood maximization, and that they are amenable to training using the Optimal Transport framework of Wasserstein Autoencoders. We find our discrete latent variable to be fully leveraged by the model when trained, without any modifications to the objective function or significant fine tuning. Our model generates comparable samples to other approaches while using relatively simple neural networks, since the discrete latent variable carries much of the descriptive burden. Furthermore, the discrete latent provides significant control over generation.

Text summarization has been a significant challenge in the Nature Process Language (NLP) field. The approach of dealing with text summarization can be roughly divided into two main paradigms: extractive and abstractive manner. The former allows capturing the most representative snippets in a document while the latter generates a summary by understanding the latent meaning in a material with a language generation model. Recently, studies found that jointly employing the extractive and abstractive summarization models can take advantage of their complementary advantages, creating both concise and informative summaries. However, the reinforced summarization models mainly depend on the ROUGE-based reward, which only has the ability to quantify the extent of word-matching rather than semantic-matching between document and summary. Meanwhile, documents are usually collected with redundant or noisy information due to the existence of repeated or irrelevant information in real-world applications. Therefore, only depending on ROUGE-based reward to optimize the reinforced summarization models may lead to biased summary generation. In this paper, we propose a novel deep \bf{Hy}brid \bf{S}ummarization with semantic weighting \bf{R}eward and latent structure \bf{D}etector (HySRD). Specifically, HySRD introduces a new reward mechanism that simultaneously takes advantage of semantic and syntactic information among documents and summaries. To effectively model the accuracy semantics, a latent structure detector is designed to incorporate the high-level latent structures in the sentence representation for information selection. Extensive experiments have been conducted on two well-known benchmark datasets \emph{CNN/Daily Mail} (short input document) and \emph{BigPatent} (long input document). The automatic evaluation shows that our approach significantly outperforms the state-of-the-art of hybrid summarization models.

Block-based programming is of great significance for cultivating children’s computational thinking. However, due to the following challenges, it is difficult to evaluate students’ programming ability in online learning systems: 1) compared with the traditional Online Judge (OJ) system, there is no standard answer for a given task in block-based programming; 2) in order to promote students’ interests, although the programs are not totally correct and unrelated to the task, the teacher will give a comparatively higher score. Therefore, current approaches involving output comparison and code analysis do not work effectively. Furthermore, deep learning methods also suffer from the problem of how to represent block code for classification. We propose a novel hybrid estimation model to address these challenges. We first learn graph embedding from the parsed Abstract Syntax Tree (AST) to present the logicality of the code. Next, we provide some methods to measure the workload and complexity of the code. Then, we extracted some key variables and task-irrelevant properties, introduced teacher bias. Finally, XGBoost was constructed for classification. Based on real-world data collected from an online Scratch platform by early-age students, our model outperforms KimCNN, ResNet-18, and Graph2Vec+XGBoost. Moreover, we provided statistical analyses and intuitive explanations to interpret the characteristics in various groups.

Assigning meaning to parts of image data is the goal of semantic image segmentation. Machine learning methods, specifically supervised learning is commonly used in a variety of tasks formulated as semantic segmentation. One of the major challenges in the supervised learning approaches is expressing and collecting the rich knowledge that experts have with respect to the meaning present in the image data. Towards this, typically a fixed set of labels is specified and experts are tasked with annotating the pixels, patches or segments in the images with the given labels. In general, however, the set of classes does not fully capture the rich semantic information present in the images. For example, in medical imaging such as histology images, the different parts of cells could be grouped and sub-grouped based on the expertise of the pathologist. To achieve such a precise semantic representation of the concepts in the image, we need access to the full depth of knowledge of the annotator. In this work, we develop a novel approach to collect segmentation annotations from experts based on psychometric testing. Our method consists of psychometric testing procedure, active query selection, query enhancement, and a deep metric learning model to achieve a patch-level image embedding that allows for semantic segmentation of images. We show the merits of our method with evaluation on the synthetically generated image, aerial image and histology image.

For lower bit-widths such as less than 8-bit, many quantization strategies include re-training in order to recover accuracy degradation. However, the re-training works against rapid deployment for wide distribution of quantized models. Therefore, post-training quantization has been getting more attention in recent years. In one example, partial quantization according to the layer sensitivity based on the accuracy after each quantization has been proposed; however, the effects of one layer quantization on the other layers has not taken into account. To further reduce the accuracy degradation, we propose a quantization scheme that considers the effects by continuously updating the accuracy after each layer quantization. Additionally, for more data compression, we extend that scheme to mixed precision, which applies a layer-by-layer fitted bit-width. Since the search space for bit allocation per layer increases exponentially with the number of layers $N$, Existing methods require computationally intensive approach such as network training. Here, we derive practical solutions to the bit allocation problem in polynomial time $O(N^2)$ using a deterministic greedy search algorithm inspired by submodular optimization without any training. For example, the proposed algorithm completes a search on ResNet18 for ImageNet in 1 hour for a single GPU. Compared to the case without updating the layer sensitivity, our method improves the accuracy of the quantized model by more than 1% with multiple convolutional neural networks. For examples, 6-bit quantization of MobileNetV2 achieves 80.1% reduction of model size with -1.10% accuracy degradation. 4-bit quantization of ResNet50 achieves 82.9% size reduction with -0.194% accuracy degradation. Furthermore, results show that the proposed method reduces the accuracy degradation by more than about 0.7% compared to various latest post-training quantization strategies.

The recent boom in the literature on entropy-regularized reinforcement learning (RL) approaches reveals that Kullback-Leibler (KL) regularization brings advantages to RL algorithms by canceling out errors under mild assumptions. However, existing analyses focus on fixed regularization with a constant weighting coefficient and do not consider cases where the coefficient is allowed to change dynamically. In this paper, we study the dynamic coefficient scheme and present the first asymptotic error bound. Based on the dynamic coefficient error bound, we propose an effective scheme to tune the coefficient according to the magnitude of error in favor of more robust learning. Complementing this development, we propose a novel algorithm, Geometric Value Iteration (GVI), that features a dynamic error-aware KL coefficient design with the aim of mitigating the impact of errors on performance. Our experiments demonstrate that GVI can effectively exploit the trade-off between learning speed and robustness over uniform averaging of a constant KL coefficient. The combination of GVI and deep networks shows stable learning behavior even in the absence of a target network, where algorithms with a constant KL coefficient would greatly oscillate or even fail to converge.

Attribution methods play a key role in generating post-hoc explanations on pre-trained models, however it has been shown that existing methods yield unfaithful and noisy explanations. In this paper, we propose a new paradigm of attribution method: we treat the model’s explanations as a part of network’s outputs then generate attribution maps from the underlying deep network. The generated attribution maps are up-sampled from the last convolutional layer of the network to obtain localization information about the target to be explained. Inspired by recent studies that showed adversarially robust models’ saliency map aligns well with human perception, we utilize attribution maps from the robust model to supervise the learned attributions. Our proposed method can produce visually plausible explanations along with the prediction in inference phase. Experiments on real datasets show that our proposed method yields more faithful explanations than post-hoc attribution methods with lighter computational costs.

Convolutional neural networks are among the most successful deep learning models used for image processing, computer vision and natural language processing applications. In this paper, we define convolution operator for numerical tabular features and thus propose feature convolutional network model for machine learning tasks. Feature convolutional networks contain feature convolution layer to extract pairwise feature convolutions in the relational feature spaces. Compared with the baseline multi-layer neural network model, the feature convolutional network gains better performance among all the experiments. The experiments results suggest that feature convolutional networks can generate efficient features automatically and provide better performance through automatic feature learning. The demo code is at https://github.com/info-ruc/FeatConvNet.

First-price auctions have largely replaced traditional bidding approaches based on Vickrey auctions in programmatic advertising. % As far as learning is concerned, first-price auctions are more challenging because the optimal bidding strategy does not only depend on the value of the item but also requires some knowledge of the other bids. % They have already given rise to several works in sequential learning, % many of which consider models for which the value of the buyer or the opponents’ maximal bid is chosen in an adversarial manner. Even in the simplest settings, this gives rise to algorithms whose pseudo-regret grows as $\sqrt{T}$ with respect to the time horizon $T$. % Focusing on the case where the buyer plays against a stationary stochastic environment, we show how to achieve significantly lower pseudo-regret: when the opponents’ maximal bid distribution is known we provide an algorithm whose pseudo-regret can be as low as $\log^2(T)$; in the case where the distribution must be learnt sequentially, a generalization of this algorithm can achieve $T^{1/3+ \epsilon}$ pseudo-regret, for any $\epsilon>0$. % To obtain these results, we introduce two novel ideas that can be of interest in their own right. First, by transposing results obtained in the posted price setting, we provide conditions under which the first-price bidding utility is locally quadratic around its optimum. Second, we leverage the observation that, on small sub-intervals, the concentration of the variations of the empirical distribution function may be controlled more accurately than by using the classical Dvoretzky-Kiefer-Wolfowitz inequality. % Numerical simulations confirm that our algorithms converge much faster than alternatives proposed in the literature for various bid distributions, including for bids collected on an actual programmatic advertising platform.

Fuzzy C-Means (FCM) clustering is a classic clustering algorithm, which is widely used in the real world. Despite the distinct advantages of FCM algorithm, whether the usage of fairness constraint in the FCM could improve clustering fairness remains fully elusive. By introducing a novel fair loss term into the objective function, a Fair Fuzzy C-Means (FFCM) algorithm was proposed in this current study. We proved that the membership value was constrained by distance and fairness in the meantime during the optimization process in the proposed objective function. By studying the Fuzzy C-Means Clustering with fairness constraint problem and proposing a fair fuzzy C-means method, this study provided mechanism understanding in achieving the fairness constraint in Fuzzy C-Means clustering and bridged up the gap of fair fuzzy clustering.

Voice cloning is the task of learning to synthesize the voice of an unseen speaker from a few samples. While current voice cloning methods achieve promising results in Text-to-Speech (TTS) synthesis for a new voice, these approaches lack the ability to control the expressiveness of synthesized audio. In this work, we propose a controllable voice cloning method that allows fine-grained control over various style aspects of the synthesized speech for an unseen speaker. We achieve this by explicitly conditioning the speech synthesis model on a speaker encoding, pitch contour and latent style tokens during training. Through both quantitative and qualitative evaluations, we show that our framework can be used for various expressive voice cloning tasks using only a few transcribed or untranscribed speech samples for a new speaker. These cloning tasks include style transfer from a reference speech, synthesizing speech directly from text, and fine-grained style control by manipulating the style conditioning variables during inference.

Cyber-Physical Systems (CPS) plays a critical role in today’s social life, especially with occasional pandemic events. With more reliance on the cyber operation of infrastructures, it is important to understand attacking mechanisms in CPS for potential solutions and defenses, where False Data Injection Attack (FDIA) is an important class. FDIA methods in the literature require the mathematical CPS model and state variable values to create an efficient attack vector, unrealistic for many attackers in the real world. Also, they do not have performance guarantee. This paper shows that it is possible to deploy a FDIA without having the CPS model and state variables information. Additionally, we prove a theoretic bound for the proposed method. Specifically, we design a scheme that learns an implicit CPS model to create tampered sensor measurements to deploy an attack based only on historical data. The proposed framework utilizes a Wasserstein generative adversarial network with two regularization terms to create such tampered measurements also known as adversarial examples. To build an attack with confidence, we present a proof based on convergence in distribution and Lipschitz norm to show that our method captures the real observed measurement distribution. This means that our model learns the complex underlying processes from the CPSs. We demonstrate the robustness and universality of our proposed framework based on two diversified adversarial examples with different systems, domains, and datasets.

We consider the expert advice problem with a low rank but noisy loss sequence, where a loss vector $l_{t} \in [-1,1]^N$ in each round $t$ is of the form $l_{t} = U v_{t} + \epsilon_{t}$ for some fixed but unknown $N \times d$ matrix $U$ called the kernel, some $d$-dimensional seed vector $v_{t} \in \mathbb{R}^{d}$, and some additional noisy term $\epsilon_t \in \mathbb{R}^{N}$ whose norm is bounded by $\epsilon$. This is a generalization of the works of Hazan et al. and Barman et al., where the former only treats noiseless loss and the latter assumes that the kernel is known in advance. In this paper, we propose an algorithm, where we re-construct the kernel under the assumptions, that the low rank loss is noised and there is no prior information about kernel. In this algorithm, we approximate the kernel by choosing a set of loss vectors with a high degree of independence from each other, and we give a regret bound of $O(d\sqrt{T}+d^{4/3}(N\epsilon)^{1/3}\sqrt{T})$. Moreover, even if in experiment, the proposed algorithm performs better than Hazan’s algorithm and Hedge algorithm.

Many practical applications suffer from the problem of imbalanced classification. The minority class has poor classification performance; on the other hand, its misclassification cost is high. One reason for classification difficulty is the intrinsic complicated distribution characteristics (CDCs) in imbalanced data itself. Classical oversampling method SMOTE generates synthetic minority class examples between neighbors, which is parameter dependent. Furthermore, due to blindness of neighbor selection, SMOTE suffers from overgeneralization in the minority class. To solve such problems, we propose an oversampling method, called extended natural neighbors based SMOTE (ExNN-SMOTE). In ExNN-SMOTE, neighbors are determined adaptively by capturing data distribution characteristics. Extensive experiments over synthetic and real datasets demonstrate the effectiveness of ExNN-SMOTE dealing with CDCs and the superiority of ExNN-SMOTE over other SMOTE-related methods.

Given a fixed parameter budget, one can build a single large neural network or create a memory-split ensemble: a pool of several smaller networks with the same total parameter count as the single network. A memory-split ensemble can outperform its single model counterpart (Lobacheva et al., 2020): a phenomenon known as the memory-split advantage (MSA). The reasons for MSA are still not yet fully understood. In particular, it is difficult in practice to predict when it will exist. This paper sheds light on the reasons underlying MSA using random feature theory. We study the dependence of the MSA on several factors: the parameter budget, the training set size, the L2 regularization and the Stochastic Gradient Descent (SGD) hyper-parameters. Using the bias-variance decomposition, we show that MSA exists when the reduction in variance due to the ensemble (\ie, \textit{ensemble gain}) exceeds the increase in squared bias due to the smaller size of the individual networks (\ie, \textit{shrinkage cost}). Taken together, our theoretical analysis demonstrates that the MSA mainly exists for the small parameter budgets relative to the training set size, and that memory-splitting can be understood as a type of regularization. Adding other forms of regularization, \eg L2 regularization, reduces the MSA. Thus, the potential benefit of memory-splitting lies primarily in the possibility of speed-up via parallel computation. Our empirical experiments with deep neural networks and large image datasets show that MSA is not a general phenomenon, but mainly exists when the number of training iterations is small.

As the damage caused by heavy rainfall is becoming more serious, the improvement of precipitation forecasts is highly demanded. For this purpose, arithmetic and Bayesian average-based methods have been proposed to integrate multiple 2D-grid forecasts. However, since a single weight is shared in the entire grid in these methods, local variations of the importance of forecasts could not be taken into account. Besides, although a variety of information is available in precipitation forecast, it would not be straightforwardly to incorporate the additional information in the existing methods. To overcome these problems, we propose an encoder-decoder-based image transformation method that generates a weight image that is optimized in a pixel-wise manner and additional information could be embedded as the channel of input images and feature maps. Through the experiment of precipitation forecast in the period from April 2018 to March 2019 in Japan, we will show that our proposed integration method outperforms existing methods.

A coreset for a set of points is a small subset of weighted points that approximately preserves important properties of the original set. Specifically, if $P$ is a set of points, $Q$ is a set of queries, and $f:P\times Q\to\mathbb{R}$ is a cost function, then a set $S\subseteq P$ with weights $w:P\to[0,\infty)$ is an $\epsilon$-coreset for some parameter $\epsilon>0$ if $\sum_{s\in S}w(s)f(s,q)$ is a $(1+\epsilon)$ multiplicative approximation to $\sum_{p\in P}f(p,q)$ for all $q\in Q$. Coresets are used to solve fundamental problems in machine learning under various big data models of computation. Many of the suggested coresets in the recent decade used, or could have used a general framework for constructing coresets whose size depends quadratically on the total sensitivity $t$. In this paper we improve this bound from $O(t^2)$ to $O(t\log t)$. Thus our results imply more space efficient solutions to a number of problems, including projective clustering, $k$-line clustering, and subspace approximation. The main technical result is a generic reduction to the sample complexity of learning a class of functions with bounded VC dimension. We show that obtaining an $(\nu,\alpha)$-sample for this class of functions with appropriate parameters $\nu$ and $\alpha$ suffices to achieve space efficient $\epsilon$-coresets. Our result implies more efficient coreset constructions for a number of interesting problems in machine learning; we show applications to $k$-median/$k$-means, $k$-line clustering, $j$-subspace approximation, and the integer $(j,k)$-projective clustering problem.

With the development of deep learning techniques, contrastive representation learning has been increasingly employed in large-scale recommender systems. For instance, deep user-item matching models can be trained by contrasting positive and negative examples and learning discriminative user and item representations. Despite their success, the distinguishable properties of the recommender system are often ignored in existing modelling. Standard methods approximate maximum likelihood estimation on user behavior data in a manner similar to language models. Specifically, the way of model optimization corresponds to approximating the user-item pointwise mutual information, which can be regarded as eliminating the influence of global item popularity on user behavior to capture intrinsic user preference. In addition, unlike the situation in language models where word frequency is relatively stable, item popularity is constantly evolving. To address these issues, we propose a novel dynamic popularity-aware (DPA) contrastive learning method for recommendation, which consists of two key components: i) a dynamic negative sampling strategy is involved to enhance the user representation, ii) a dynamic prediction recovery is adopted by the real-time item popularity. The proposed strategy can be naturally overlaid on any contrastive learning-based matching model to more accurately capture user interest and system dynamics. Finally, the effectiveness of the proposed strategy is demonstrated through comprehensive experiments on an e-commerce scenario of Alibaba Group.

This paper introduces Dynamic $Q$-value Coordination Graph (QCGraph) for cooperative multi-agent reinforcement learning. QCGraph aims to dynamically represent and generalize through factorizing the joint value function of all agents according to dynamically created coordination graph based on subsets of agents. The value can be maximized by message passing at both a local and global level along the graph which allows training the value function end-to-end. The coordination graph is dynamically generated and used to generate the payoff functions which are approximated using graph neural networks and parameter sharing to improve generalization over the state-action space. We show that QCGraph can solve a variety of challenging multi-agent tasks being superior to other value factorization approaches.