arXiv subjects

Vincent Y. F. Tan

Publications and source records attributed to Vincent Y. F. Tan.

At least 127 records · Page 7Linked to original sources

arXiv2021

Risk-Constrained Thompson Sampling for CVaR Bandits

The multi-armed bandit (MAB) problem is a ubiquitous decision-making problem that exemplifies the exploration-exploitation tradeoff. Standard formulations exclude risk in decision making. Risk notably complicates the basic reward-maximising objective, in part because there is no universally agreed definition of it. In this paper, we consider a popular risk measure in quantitative finance known as the Conditional Value at Risk (CVaR). We explore the performance of a Thompson Sampling-based algorithm CVaR-TS under this risk measure. We provide comprehensive comparisons between our regret bounds with state-of-the-art L/UCB-based algorithms in comparable settings and demonstrate their clear improvement in performance. We also include numerical simulations to empirically verify that CVaR-TS outperforms other L/UCB-based algorithms.

cs.LG↗

arXiv2021

On Non-Interactive Simulation of Binary Random Variables

We leverage proof techniques Fourier analysis and an existing result in coding theory to derive new bounds for the problem of non-interactive simulation of binary random variables. Previous bounds in the literature were derived by applying data processing inequalities concerning maximal correlation or hypercontractivity. We show that our bounds are sharp in some regimes. For a specific instance of problem parameters, our main result answers an open problem posed by E. Mossel in 2017. As by-products of our analyses, various new properties of the average distance and distance enumerator of binary block codes are established.

cs.IT↗

arXiv2021

SGA: A Robust Algorithm for Partial Recovery of Tree-Structured Graphical Models with Noisy Samples

We consider learning Ising tree models when the observations from the nodes are corrupted by independent but non-identically distributed noise with unknown statistics. Katiyar et al. (2020) showed that although the exact tree structure cannot be recovered, one can recover a partial tree structure; that is, a structure belonging to the equivalence class containing the true tree. This paper presents a systematic improvement of Katiyar et al. (2020). First, we present a novel impossibility result by deriving a bound on the necessary number of samples for partial recovery. Second, we derive a significantly improved sample complexity result in which the dependence on the minimum correlation $ρ_{\min}$ is $ρ_{\min}^{-8}$ instead of $ρ_{\min}^{-24}$. Finally, we propose Symmetrized Geometric Averaging (SGA), a more statistically robust algorithm for partial tree recovery. We provide error exponent analyses and extensive numerical results on a variety of trees to show that the sample complexity of SGA is significantly better than the algorithm of Katiyar et al. (2020). SGA can be readily extended to Gaussian models and is shown via numerical experiments to be similarly superior.

stat.ML↗

arXiv2021

An Interpretable Intensive Care Unit Mortality Risk Calculator

Mortality risk is a major concern to patients have just been discharged from the intensive care unit (ICU). Many studies have been directed to construct machine learning models to predict such risk. Although these models are highly accurate, they are less amenable to interpretation and clinicians are typically unable to gain further insights into the patients' health conditions and the underlying factors that influence their mortality risk. In this paper, we use patients' profiles extracted from the MIMIC-III clinical database to construct risk calculators based on different machine learning techniques such as logistic regression, decision trees, random forests and multilayer perceptrons. We perform an extensive benchmarking study that compares the most salient features as predicted by various methods. We observe a high degree of agreement across the considered machine learning methods; in particular, the cardiac surgery recovery unit, age, and blood urea nitrogen levels are commonly predicted to be the most salient features for determining patients' mortality risks. Our work has the potential for clinicians to interpret risk predictions.

stat.AP↗

arXiv2021

Community Detection and Matrix Completion with Social and Item Similarity Graphs

We consider the problem of recovering a binary rating matrix as well as clusters of users and items based on a partially observed matrix together with side-information in the form of social and item similarity graphs. These two graphs are both generated according to the celebrated stochastic block model (SBM). We develop lower and upper bounds on sample complexity that match for various scenarios. Our information-theoretic results quantify the benefits of the availability of the social and item similarity graphs. Further analysis reveals that under certain scenarios, the social and item similarity graphs produce an interesting synergistic effect. This means that observing two graphs is strictly better than observing just one in terms of reducing the sample complexity.

cs.IT↗

arXiv2020

State Masking Over a Two-State Compound Channel

We consider fundamental limits for communicating over a compound channel when the state of the channel needs to be masked. Our model is closely related to an area of study known as covert communication that is a setting in which the transmitter wishes to communicate to legitimate receiver(s) while ensuring that the communication is not detected by an adversary. The main contribution in our two-state masking setup is the establishment of bounds on the throughput-key region when the constraint that quantifies how much the states are masked is defined to be the total variation distance between the two channel-induced distributions. For the scenario in which the key length is infinite, we provide sufficient conditions for when the bounds to coincide for the throughput, which follows the square-root law. Numerical examples, including that of a Gaussian channel, are provided to illustrate our results.

cs.IT↗

arXiv2020

RAIN: A Simple Approach for Robust and Accurate Image Classification Networks

It has been shown that the majority of existing adversarial defense methods achieve robustness at the cost of sacrificing prediction accuracy. The undesirable severe drop in accuracy adversely affects the reliability of machine learning algorithms and prohibits their deployment in realistic applications. This paper aims to address this dilemma by proposing a novel preprocessing framework, which we term Robust and Accurate Image classificatioN(RAIN), to improve the robustness of given CNN classifiers and, at the same time, preserve their high prediction accuracies. RAIN introduces a new randomization-enhancement scheme. It applies randomization over inputs to break the ties between the model forward prediction path and the backward gradient path, thus improving the model robustness. However, similar to existing preprocessing-based methods, the randomized process will degrade the prediction accuracy. To understand why this is the case, we compare the difference between original and processed images, and find it is the loss of high-frequency components in the input image that leads to accuracy drop of the classifier. Based on this finding, RAIN enhances the input's high-frequency details to retain the CNN's high prediction accuracy. Concretely, RAIN consists of two novel randomization modules: randomized small circular shift (RdmSCS) and randomized down-upsampling (RdmDU). The RdmDU module randomly downsamples the input image, and then the RdmSCS module circularly shifts the input image along a randomly chosen direction by a small but random number of pixels. Finally, the RdmDU module performs upsampling with a detail-enhancement model, such as deep super-resolution networks. We conduct extensive experiments on the STL10 and ImageNet datasets to verify the effectiveness of RAIN against various types of adversarial attacks.

cs.CR↗

arXiv2020

Variable-Length Source Dispersions Differ under Maximum and Average Error Criteria

Variable-length compression without prefix-free constraints and with side-information available at both encoder and decoder is considered. Instead of requiring the code to be error-free, we allow for it to have a non-vanishing error probability. We derive one-shot bounds on the optimal average codeword length by proposing two new information quantities; namely, the conditional and unconditional $\varepsilon$-cutoff entropies. Using these one-shot bounds, we obtain the second-order asymptotics of the problem under two different formalisms---the average and maximum probabilities of error over the realization of the side-information. While the first-order terms in the asymptotic expansions for both formalisms are identical, we find that the source dispersion under the average error formalism is, in most cases, strictly smaller than its maximum error counterpart. Applications to a certain class of guessing problems, previously studied by Kuzuoka [\emph{{IEEE} Trans.\ Inf.\ Theory}, vol.~66, no.~3, pp.~1674--1690, 2020], are also discussed.

cs.IT↗

arXiv2020

On the Error Exponent of Approximate Sufficient Statistics for M-ary Hypothesis Testing

Consider the problem of detecting one of M i.i.d. Gaussian signals corrupted in white Gaussian noise. Conventionally, matched filters are used for detection. We first show that the outputs of the matched filter form a set of asymptotically optimal sufficient statistics in the sense of maximizing the error exponent of detecting the true signal. In practice, however, M may be large which motivates the design and analysis of a reduced set of N statistics which we term approximate sufficient statistics. Our construction of these statistics is based on a small set of filters that project the outputs of the matched filters onto a lower-dimensional vector using a sensing matrix. We consider a sequence of sensing matrices that has the desiderata of row orthonormality and low coherence. We analyze the performance of the resulting maximum likelihood (ML) detector, which leads to an achievable bound on the error exponent based on the approximate sufficient statistics; this bound recovers the original error exponent when N = M. We compare this to a bound that we obtain by analyzing a modified form of the Reduced Dimensionality Detector (RDD) proposed by Xie, Eldar, and Goldsmith [IEEE Trans. on Inform. Th., 59(6):3858-3874, 2013]. We show that by setting the sensing matrices to be column-normalized group Hadamard matrices, the exponents derived are ensemble-tight, i.e., our analysis is tight on the exponential scale given the sensing matrices and the decoding rule. Finally, we derive some properties of the exponents, showing, in particular, that they increase linearly in the compression ratio N/M.

cs.IT↗

arXiv2020

Thompson Sampling Algorithms for Mean-Variance Bandits

The multi-armed bandit (MAB) problem is a classical learning task that exemplifies the exploration-exploitation tradeoff. However, standard formulations do not take into account {\em risk}. In online decision making systems, risk is a primary concern. In this regard, the mean-variance risk measure is one of the most common objective functions. Existing algorithms for mean-variance optimization in the context of MAB problems have unrealistic assumptions on the reward distributions. We develop Thompson Sampling-style algorithms for mean-variance MAB and provide comprehensive regret analyses for Gaussian and Bernoulli bandits with fewer assumptions. Our algorithms achieve the best known regret bounds for mean-variance MABs and also attain the information-theoretic bounds in some parameter regimes. Empirical simulations show that our algorithms significantly outperform existing LCB-based algorithms for all risk tolerances.

cs.LG↗

arXiv2020

Second-Order Asymptotics of Sequential Hypothesis Testing

We consider the classical sequential binary hypothesis testing problem in which there are two hypotheses governed respectively by distributions $P_0$ and $P_1$ and we would like to decide which hypothesis is true using a sequential test. It is known from the work of Wald and Wolfowitz that as the expectation of the length of the test grows, the optimal type-I and type-II error exponents approach the relative entropies $D(P_1\|P_0)$ and $D(P_0\|P_1)$. We refine this result by considering the optimal backoff---or second-order asymptotics---from the corner point of the achievable exponent region $(D(P_1\|P_0),D(P_0\|P_1))$ under two different constraints on the length of the test (or the sample size). First, we consider a probabilistic constraint in which the probability that the length of test exceeds a prescribed integer $n$ is less than a certain threshold $0<\varepsilon <1$. Second, the expectation of the sample size is bounded by $n$. In both cases, and under mild conditions, the second-order asymptotics is characterized exactly. Numerical examples are provided to illustrate our results.

cs.IT↗

arXiv2020

Best Arm Identification for Cascading Bandits in the Fixed Confidence Setting

We design and analyze CascadeBAI, an algorithm for finding the best set of $K$ items, also called an arm, within the framework of cascading bandits. An upper bound on the time complexity of CascadeBAI is derived by overcoming a crucial analytical challenge, namely, that of probabilistically estimating the amount of available feedback at each step. To do so, we define a new class of random variables (r.v.'s) which we term as left-sided sub-Gaussian r.v.'s; these are r.v.'s whose cumulant generating functions (CGFs) can be bounded by a quadratic only for non-positive arguments of the CGFs. This enables the application of a sufficiently tight Bernstein-type concentration inequality. We show, through the derivation of a lower bound on the time complexity, that the performance of CascadeBAI is optimal in some practical regimes. Finally, extensive numerical simulations corroborate the efficacy of CascadeBAI as well as the tightness of our upper bound on its time complexity.

cs.LG↗

arXiv2020

Unsupervised Image Noise Modeling with Self-Consistent GAN

Noise modeling lies in the heart of many image processing tasks. However, existing deep learning methods for noise modeling generally require clean and noisy image pairs for model training; these image pairs are difficult to obtain in many realistic scenarios. To ameliorate this problem, we propose a self-consistent GAN (SCGAN), that can directly extract noise maps from noisy images, thus enabling unsupervised noise modeling. In particular, the SCGAN introduces three novel self-consistent constraints that are complementary to one another, viz.: the noise model should produce a zero response over a clean input; the noise model should return the same output when fed with a specific pure noise input; and the noise model also should re-extract a pure noise map if the map is added to a clean image. These three constraints are simple yet effective. They jointly facilitate unsupervised learning of a noise model for various noise types. To demonstrate its wide applicability, we deploy the SCGAN on three image processing tasks including blind image denoising, rain streak removal, and noisy image super-resolution. The results demonstrate the effectiveness and superiority of our method over the state-of-the-art methods on a variety of benchmark datasets, even though the noise types vary significantly and paired clean images are not available.

cs.CV↗

arXiv2020

Exact Asymptotics for Learning Tree-Structured Graphical Models with Side Information: Noiseless and Noisy Samples

Given side information that an Ising tree-structured graphical model is homogeneous and has no external field, we derive the exact asymptotics of learning its structure from independently drawn samples. Our results, which leverage the use of probabilistic tools from the theory of strong large deviations, refine the large deviation (error exponents) results of Tan, Anandkumar, Tong, and Willsky [IEEE Trans. on Inform. Th., 57(3):1714--1735, 2011] and strictly improve those of Bresler and Karzand [Ann. Statist., 2020]. In addition, we extend our results to the scenario in which the samples are observed in random noise. In this case, we show that they strictly improve on the recent results of Nikolakakis, Kalogerias, and Sarwate [Proc. AISTATS, 1771--1782, 2019]. Our theoretical results demonstrate keen agreement with experimental results for sample sizes as small as that in the hundreds.

stat.ML↗

arXiv2020

Second- and Third-Order Asymptotics of the Continuous-Time Poisson Channel

The paper derives the optimal second-order coding rate for the continuous-time Poisson channel. We also obtain bounds on the third-order coding rate. This is the first instance of a second-order result for a continuous-time channel. The converse proof hinges on a novel construction of an output distribution induced by Wyner's discretized channel and the construction of an appropriate $ε$-net of the input probability simplex. While the achievability proof follows the general program to prove the third-order term for non-singular discrete memoryless channels put forth by Polyanskiy, several non-standard techniques -- such as new definitions and bounds on the probabilities of typical sets using logarithmic Sobolev inequalities -- are employed to handle the continuous nature of the channel.

cs.IT↗

arXiv2020

Asymptotic Expansions of Smooth Rényi Entropies and Their Applications

This study considers the unconditional smooth Rényi entropy, the smooth conditional Rényi entropy proposed by Kuzuoka [\emph{IEEE Trans.\ Inf.\ Theory}, vol.~66, no.~3, pp.~1674--1690, 2020], and a new quantity which we term the conditional smooth Rényi entropy. In particular, we examine asymptotic expansions of these entropies when the underlying source with its side-information is stationary and memoryless. Using these smooth Rényi entropies, we establish one-shot coding theorems of several information-theoretic problems: Campbell's source coding, guessing problems, and task encoding problems, all allowing errors. In each problem, we consider two error formalisms: the average and maximum error criteria, where the averaging and maximization are taken with respect to the side-information of the source. Applying our asymptotic expansions to the derived one-shot coding theorems, we derive various asymptotic fundamental limits for these problems when their error probabilities are allowed to be non-vanishing. We show that, in non-degenerate settings, the first-order fundamental limits differ under the average and maximum error criteria. This is in contrast to a different but related setting considered by the present authors (for variable-length conditional source coding allowing errors) in which the first-order terms are identical but the second-order terms are different under these criteria.

cs.IT↗

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↗