arXiv subjects

Lei Yu

Publications and source records attributed to Lei Yu.

At least 235 records · Page 13Linked to original sources

arXiv2020

Intra-Ensemble in Neural Networks

Improving model performance is always the key problem in machine learning including deep learning. However, stand-alone neural networks always suffer from marginal effect when stacking more layers. At the same time, ensemble is an useful technique to further enhance model performance. Nevertheless, training several independent deep neural networks for ensemble costs multiple resources. If so, is it possible to utilize ensemble in only one neural network? In this work, we propose Intra-Ensemble, an end-to-end ensemble strategy with stochastic channel recombination operations to train several sub-networks simultaneously within one neural network. Additional parameter size is marginal since the majority of parameters are mutually shared. Meanwhile, stochastic channel recombination significantly increases the diversity of sub-networks, which finally enhances ensemble performance. Extensive experiments and ablation studies prove the applicability of intra-ensemble on various kinds of datasets and network architectures.

cs.CV↗

arXiv2020

Robust Time-Frequency Reconstruction by Learning Structured Sparsity

Time-frequency distributions (TFDs) play a vital role in providing descriptive analysis of non-stationary signals involved in realistic scenarios. It is well known that low time-frequency (TF) resolution and the emergency of cross-terms (CTs) are two main issues, which make it difficult to analyze and interpret practical signals using TFDs. In order to address these issues, we propose the U-Net aided iterative shrinkage-thresholding algorithm (U-ISTA) for reconstructing a near-ideal TFD by exploiting structured sparsity in signal TF domain. Specifically, the signal ambiguity function is firstly compressed, followed by unfolding the ISTA as a recurrent neural network. To consider continuously distributed characteristics of signals, a structured sparsity constraint is incorporated into the unfolded ISTA by regarding the U-Net as an adaptive threshold block, in which structure-aware thresholds are learned from enormous training data to exploit the underlying dependencies among neighboring TF coefficients. The proposed U-ISTA model is trained by both non-overlapped and overlapped synthetic signals including closely and far located non-stationary components. Experimental results demonstrate that the robust U-ISTA achieves superior performance compared with state-of-the-art algorithms, and gains a high TF resolution with CTs greatly eliminated even in low signal-to-noise ratio (SNR) environments.

eess.SP↗

arXiv2020

Capturing document context inside sentence-level neural machine translation models with self-training

Neural machine translation (NMT) has arguably achieved human level parity when trained and evaluated at the sentence-level. Document-level neural machine translation has received less attention and lags behind its sentence-level counterpart. The majority of the proposed document-level approaches investigate ways of conditioning the model on several source or target sentences to capture document context. These approaches require training a specialized NMT model from scratch on parallel document-level corpora. We propose an approach that doesn't require training a specialized model on parallel document-level corpora and is applied to a trained sentence-level NMT model at decoding time. We process the document from left to right multiple times and self-train the sentence-level model on pairs of source sentences and generated translations. Our approach reinforces the choices made by the model, thus making it more likely that the same choices will be made in other sentences in the document. We evaluate our approach on three document-level datasets: NIST Chinese-English, WMT'19 Chinese-English and OpenSubtitles English-Russian. We demonstrate that our approach has higher BLEU score and higher human preference than the baseline. Qualitative analysis of our approach shows that choices made by model are consistent across the document.

cs.CL↗

arXiv2020

Matching Neuromorphic Events and Color Images via Adversarial Learning

The event camera has appealing properties: high dynamic range, low latency, low power consumption and low memory usage, and thus provides complementariness to conventional frame-based cameras. It only captures the dynamics of a scene and is able to capture almost "continuous" motion. However, different from frame-based camera that reflects the whole appearance as scenes are, the event camera casts away the detailed characteristics of objects, such as texture and color. To take advantages of both modalities, the event camera and frame-based camera are combined together for various machine vision tasks. Then the cross-modal matching between neuromorphic events and color images plays a vital and essential role. In this paper, we propose the Event-Based Image Retrieval (EBIR) problem to exploit the cross-modal matching task. Given an event stream depicting a particular object as query, the aim is to retrieve color images containing the same object. This problem is challenging because there exists a large modality gap between neuromorphic events and color images. We address the EBIR problem by proposing neuromorphic Events-Color image Feature Learning (ECFL). Particularly, the adversarial learning is employed to jointly model neuromorphic events and color images into a common embedding space. We also contribute to the community N-UKbench and EC180 dataset to promote the development of EBIR problem. Extensive experiments on our datasets show that the proposed method is superior in learning effective modality-invariant representation to link two different modalities.

cs.CV↗

arXiv2020

On Exact and $\infty$-Rényi Common Informations

Recently, two extensions of Wyner's common information\textemdash exact and Rényi common informations\textemdash were introduced respectively by Kumar, Li, and El Gamal (KLE), and the present authors. The class of common information problems involves determining the minimum rate of the common input to two independent processors needed to exactly or approximately generate a target joint distribution. For the exact common information problem, exact generation of the target distribution is required, while for Wyner's and $α$-Rényi common informations, the relative entropy and Rényi divergence with order $α$ were respectively used to quantify the discrepancy between the synthesized and target distributions. The exact common information is larger than or equal to Wyner's common information. However, it was hitherto unknown whether the former is strictly larger than the latter for some joint distributions. In this paper, we first establish the equivalence between the exact and $\infty$-Rényi common informations, and then provide single-letter upper and lower bounds for these two quantities. For doubly symmetric binary sources, we show that the upper and lower bounds coincide, which implies that for such sources, the exact and $\infty$-Rényi common informations are completely characterized. Interestingly, we observe that for such sources, these two common informations are strictly larger than Wyner's. This answers an open problem posed by KLE. Furthermore, we extend Wyner's, $\infty$-Rényi, and exact common informations to sources with countably infinite or continuous alphabets, including Gaussian sources.

cs.IT↗

arXiv2020

Corrections to "Wyner's Common Information under Rényi Divergence Measures"

In this correspondence, we correct an erroneous result on the achievability part of the Rényi common information with order $1+s\in(1,2]$ in [1]. The new achievability result (upper bound) of the Rényi common information no longer coincides with Wyner's common information. We also provide a new converse result (lower bound) in this correspondence for the Rényi common information with order $1+s\in(1,\infty]$. Numerical results show that for doubly symmetric binary sources, the new upper and lower bounds coincide for the order $1+s\in(1,2]$ and they are both strictly larger than Wyner's common information for this case.

cs.IT↗

arXiv2020

Exact minimum codegree thresholds for $K_4^-$-covering and $K_5^-$-covering

Given two $3$-graphs $F$ and $H$, an $F$-covering of $H$ is a collection of copies of $F$ in $H$ such that each vertex of $H$ is contained in at least one copy of them. Let {$c_2(n,F)$} be the maximum integer $t$ such that every 3-graph with minimum codegree greater than $t$ has an $F$-covering. In this note, we answer an open problem of Falgas-Ravry and Zhao (SIAM J. Discrete Math., 2016) by determining the exact value of {$c_2(n, K_4^-)$} and {$c_2(n, K_5^-)$}, where $K_t^-$ is the complete $3$-graph on $t$ vertices with one edge removed.

math.CO↗

arXiv2019

Differentially Private Model Publishing for Deep Learning

Deep learning techniques based on neural networks have shown significant success in a wide range of AI tasks. Large-scale training datasets are one of the critical factors for their success. However, when the training datasets are crowdsourced from individuals and contain sensitive information, the model parameters may encode private information and bear the risks of privacy leakage. The recent growing trend of the sharing and publishing of pre-trained models further aggravates such privacy risks. To tackle this problem, we propose a differentially private approach for training neural networks. Our approach includes several new techniques for optimizing both privacy loss and model accuracy. We employ a generalization of differential privacy called concentrated differential privacy(CDP), with both a formal and refined privacy loss analysis on two different data batching methods. We implement a dynamic privacy budget allocator over the course of training to improve model accuracy. Extensive experiments demonstrate that our approach effectively improves privacy loss accounting, training efficiency and model quality under a given privacy budget.

cs.CR↗

arXiv2019

Wyner's Common Information under Rényi Divergence Measures

We study a generalized version of Wyner's common information problem (also coined the distributed source simulation problem). The original common information problem consists in understanding the minimum rate of the common input to independent processors to generate an approximation of a joint distribution when the distance measure used to quantify the discrepancy between the synthesized and target distributions is the normalized relative entropy. Our generalization involves changing the distance measure to the unnormalized and normalized Rényi divergences of order $α=1+s\in[0,2]$. We show that the minimum rate needed to ensure the Rényi divergences between the distribution induced by a code and the target distribution vanishes remains the same as the one in Wyner's setting, except when the order $α=1+s=0$. This implies that Wyner's common information is rather robust to the choice of distance measure employed. As a byproduct of the proofs used to the establish the above results, the exponential strong converse for the common information problem under the total variation distance measure is established.

cs.IT↗

arXiv2019

A Mutual Information Maximization Perspective of Language Representation Learning

We show state-of-the-art word representation learning methods maximize an objective function that is a lower bound on the mutual information between different parts of a word sequence (i.e., a sentence). Our formulation provides an alternative perspective that unifies classical word embedding models (e.g., Skip-gram) and modern contextual embeddings (e.g., BERT, XLNet). In addition to enhancing our theoretical understanding of these methods, our derivation leads to a principled framework that can be used to construct new self-supervised tasks. We provide an example by drawing inspirations from related methods based on mutual information maximization that have been successful in computer vision, and introduce a simple self-supervised objective that maximizes the mutual information between a global sentence representation and n-grams in the sentence. Our analysis offers a holistic view of representation learning methods to transfer knowledge and translate progress across multiple domains (e.g., natural language processing, computer vision, audio processing).

cs.CL↗

arXiv2019

Effects of Differential Privacy and Data Skewness on Membership Inference Vulnerability

Membership inference attacks seek to infer the membership of individual training instances of a privately trained model. This paper presents a membership privacy analysis and evaluation system, called MPLens, with three unique contributions. First, through MPLens, we demonstrate how membership inference attack methods can be leveraged in adversarial machine learning. Second, through MPLens, we highlight how the vulnerability of pre-trained models under membership inference attack is not uniform across all classes, particularly when the training data itself is skewed. We show that risk from membership inference attacks is routinely increased when models use skewed training data. Finally, we investigate the effectiveness of differential privacy as a mitigation technique against membership inference attacks. We discuss the trade-offs of implementing such a mitigation strategy with respect to the model complexity, the learning task complexity, the dataset complexity and the privacy parameter settings. Our empirical results reveal that (1) minority groups within skewed datasets display increased risk for membership inference and (2) differential privacy presents many challenging trade-offs as a mitigation technique to membership inference risk.

cs.CR↗

arXiv2019

Maximum Entropy Diverse Exploration: Disentangling Maximum Entropy Reinforcement Learning

Two hitherto disconnected threads of research, diverse exploration (DE) and maximum entropy RL have addressed a wide range of problems facing reinforcement learning algorithms via ostensibly distinct mechanisms. In this work, we identify a connection between these two approaches. First, a discriminator-based diversity objective is put forward and connected to commonly used divergence measures. We then extend this objective to the maximum entropy framework and propose an algorithm Maximum Entropy Diverse Exploration (MEDE) which provides a principled method to learn diverse behaviors. A theoretical investigation shows that the set of policies learned by MEDE capture the same modalities as the optimal maximum entropy policy. In effect, the proposed algorithm disentangles the maximum entropy policy into its diverse, constituent policies. Experiments show that MEDE is superior to the state of the art in learning high performing and diverse policies.

cs.LG↗

arXiv2019

Demystifying Learning Rate Policies for High Accuracy Training of Deep Neural Networks

Learning Rate (LR) is an important hyper-parameter to tune for effective training of deep neural networks (DNNs). Even for the baseline of a constant learning rate, it is non-trivial to choose a good constant value for training a DNN. Dynamic learning rates involve multi-step tuning of LR values at various stages of the training process and offer high accuracy and fast convergence. However, they are much harder to tune. In this paper, we present a comprehensive study of 13 learning rate functions and their associated LR policies by examining their range parameters, step parameters, and value update parameters. We propose a set of metrics for evaluating and selecting LR policies, including the classification confidence, variance, cost, and robustness, and implement them in LRBench, an LR benchmarking system. LRBench can assist end-users and DNN developers to select good LR policies and avoid bad LR policies for training their DNNs. We tested LRBench on Caffe, an open source deep learning framework, to showcase the tuning optimization of LR policies. Evaluated through extensive experiments, we attempt to demystify the tuning of LR policies by identifying good LR policies with effective LR value ranges and step sizes for LR update schedules.

cs.LG↗

arXiv2019

Deep learning scheme for recovery of broadband microwave photonic receiving systems in transceivers without expert knowledge and system priors

In regular microwave photonic (MWP) receiving systems, broadband signals are processed in the analog domain before they are transformed to the digital domain for further processing and storage. However, the quality of the signals may be degraded by defective photonic analog links, especially in a complicated MWP system. Here, we show a unified deep learning scheme that recovers the distorted broadband signals as they are transformed to the digital domain. The neural network could automatically learn the end-to-end inverse responses of the distortion effects of actual photonic analog links from data without expert knowledge and system priors. Hence, by shifting or augmenting the datasets, the neural network is potential to be generalized to various MWP receiving systems. We conduct experiments by nontrivial MWP systems with complicated waveforms. Results validate the effectiveness, general applicability and the noise-robustness of the proposed scheme, showing its superior performance in practical MWP systems. Therefore, the proposed deep learning scheme facilitates the low-cost performance improvement of MWP receiving systems, as well as the next-generation broadband transceivers, including radars, communications, and microwave imaging.

eess.SP↗

arXiv2019

An Improved Linear Programming Bound on the Average Distance of a Binary Code

Ahlswede and Katona (1977) posed the following isodiametric problem in Hamming spaces: For every $n$ and $1\le M\le2^{n}$, determine the minimum average Hamming distance of binary codes with length $n$ and size $M$. Fu, Wei, and Yeung (2001) used linear programming duality to derive a lower bound on the minimum average distance. However, their linear programming approach was not completely exploited. In this paper, we improve Fu-Wei-Yeung's bound by finding a better feasible solution to their dual program. For fixed $0<a\le1/2$ and for $M=\left\lceil a2^{n}\right\rceil $, our feasible solution attains the asymptotically optimal value of Fu-Wei-Yeung's dual program as $n\to\infty$. Hence for $0<a\le1/2$, all possible asymptotic bounds that can be derived by Fu-Wei-Yeung's linear program have been characterized. Furthermore, noting that the average distance of a code is closely related to weights of Fourier coefficients of a Boolean function, we also apply the linear programming technique to prove bounds on Fourier weights of a Boolean function of various degrees.

math.CO↗

arXiv2019

Learning Cluster Structured Sparsity by Reweighting

Recently, the paradigm of unfolding iterative algorithms into finite-length feed-forward neural networks has achieved a great success in the area of sparse recovery. Benefit from available training data, the learned networks have achieved state-of-the-art performance in respect of both speed and accuracy. However, the structure behind sparsity, imposing constraint on the support of sparse signals, is often an essential prior knowledge but seldom considered in the existing networks. In this paper, we aim at bridging this gap. Specifically, exploiting the iterative reweighted $\ell_1$ minimization (IRL1) algorithm, we propose to learn the cluster structured sparsity (CSS) by rewegihting adaptively. In particular, we first unfold the Reweighted Iterative Shrinkage Algorithm (RwISTA) into an end-to-end trainable deep architecture termed as RW-LISTA. Then instead of the element-wise reweighting, the global and local reweighting manner are proposed for the cluster structured sparse learning. Numerical experiments further show the superiority of our algorithm against both classical algorithms and learning-based networks on different tasks.

cs.LG↗

arXiv2019

Exact Channel Synthesis

We consider the exact channel synthesis problem. This problem concerns the determination of the minimum amount of information required to create exact correlation remotely when there is a certain rate of randomness shared by two terminals. This problem generalizes an existing approximate version, in which the generated joint distribution is required to be close to a target distribution under the total variation (TV) distance measure (instead being exactly equal to the target distribution). We provide single-letter inner and outer bounds on the admissible region of the shared randomness rate and the communication rate for the exact channel synthesis problem. These two bounds coincide for doubly symmetric binary sources. We observe that for such sources, the admissible rate region for exact channel synthesis is strictly included in that for the TV-approximate version. We also extend the exact and TV-approximate channel synthesis problems to sources with countably infinite alphabets and continuous sources; the latter includes Gaussian sources. As by-products, lemmas concerning soft-covering under Rényi divergence measures are derived.

cs.IT↗

arXiv2019

Unsupervised Recurrent Neural Network Grammars

Recurrent neural network grammars (RNNG) are generative models of language which jointly model syntax and surface structure by incrementally generating a syntax tree and sentence in a top-down, left-to-right order. Supervised RNNGs achieve strong language modeling and parsing performance, but require an annotated corpus of parse trees. In this work, we experiment with unsupervised learning of RNNGs. Since directly marginalizing over the space of latent trees is intractable, we instead apply amortized variational inference. To maximize the evidence lower bound, we develop an inference network parameterized as a neural CRF constituency parser. On language modeling, unsupervised RNNGs perform as well their supervised counterparts on benchmarks in English and Chinese. On constituency grammar induction, they are competitive with recent neural language models that induce tree structures from words through attention mechanisms.

cs.CL↗