Search

We analyze the performance of the least absolute shrinkage and selection operator (Lasso) for the linear model when the number of regressors $N$ grows larger keeping the true support size $d$ finite, i.e., the ultra-sparse case. The result is based on a novel treatment of the non-rigorous replica method in statistical physics, which has been applied only to problem settings where $N$ ,$d$ and the number of observations $M$ tend to infinity at the same rate. Our analysis makes it possible to assess the average performance of Lasso with Gaussian sensing matrices without assumptions on the scaling of $N$ and $M$, the noise distribution, and the profile of the true signal. Under mild conditions on the noise distribution, the analysis also offers a lower bound on the sample complexity necessary for partial and perfect support recovery when $M$ diverges as $M = O(\log N)$. The obtained bound for perfect support recovery is a generalization of that given in previous literature, which only considers the case of Gaussian noise and diverging $d$. Extensive numerical experiments strongly support our analysis., To appear in AISTATS 2023

We theoretically analyze the typical learning performance of $\ell_{1}$-regularized linear regression ($\ell_1$-LinR) for Ising model selection using the replica method from statistical mechanics. For typical random regular graphs in the paramagnetic phase, an accurate estimate of the typical sample complexity of $\ell_1$-LinR is obtained. Remarkably, despite the model misspecification, $\ell_1$-LinR is model selection consistent with the same order of sample complexity as $\ell_{1}$-regularized logistic regression ($\ell_1$-LogR), i.e., $M=\mathcal{O}\left(\log N\right)$, where $N$ is the number of variables of the Ising model. Moreover, we provide an efficient method to accurately predict the non-asymptotic behavior of $\ell_1$-LinR for moderate $M, N$, such as precision and recall. Simulations show a fairly good agreement between theoretical predictions and experimental results, even for graphs with many loops, which supports our findings. Although this paper mainly focuses on $\ell_1$-LinR, our method is readily applicable for precisely characterizing the typical learning performances of a wide class of $\ell_{1}$-regularized $M$-estimators including $\ell_1$-LogR and interaction screening., Comment: Accepted to NeurIPS 2021. Camera-ready version with supplementary materials

Resampling techniques are widely used in statistical inference and ensemble learning, in which estimators' statistical properties are essential. However, existing methods are computationally demanding, because repetitions of estimation/learning via numerical optimization/integral for each resampled data are required. In this study, we introduce a computationally efficient method to resolve such problem: replicated vector approximate message passing. This is based on a combination of the replica method of statistical physics and an accurate approximate inference algorithm, namely the vector approximate message passing of information theory. The method provides tractable densities without repeating estimation/learning, and the densities approximately offer an arbitrary degree of the estimators' moment in practical time. In the experiment, we apply the proposed method to the stability selection method, which is commonly used in variable selection problems. The numerical results show its fast convergence and high approximation accuracy for problems involving both synthetic and real-world datasets., Comment: 10 pages, 3 figures

We consider the statistical properties of interaction parameter estimates obtained by the direct coupling analysis (DCA) approach to learning interactions from large data sets. Assuming that the data are generated from a random background distribution, we determine the distribution of inferred interactions. Two inference methods are considered: the L2 regularized naive mean-field inference procedure (regularized least squares, RLS), and the pseudo-likelihood maximization (plmDCA). For RLS we also study a model where the data matrix elements are real numbers, identically and independently generated from a Gaussian distribution; in this setting we analytically find that the distribution of the inferred interactions is Gaussian. For data of Boolean type, more realistic in practice, the inferred interactions do not generally follow a Gaussian. However, extensive numerical simulations indicate that their distribution can be characterized by a single function determined by a few system parameters after normalization by the standard deviation. This property holds for both RLS and plmDCA and may be exploitable for inferring the distribution of extremely large interactions from simulations for smaller system sizes., 5 pages, 5 figures

We develop an approximate formula for evaluating a cross-validation estimator of predictive likelihood for multinomial logistic regression regularized by an $\ell_1$-norm. This allows us to avoid repeated optimizations required for literally conducting cross-validation; hence, the computational time can be significantly reduced. The formula is derived through a perturbative approach employing the largeness of the data size and the model dimensionality. An extension to the elastic net regularization is also addressed. The usefulness of the approximate formula is demonstrated on simulated data and the ISOLET dataset from the UCI machine learning repository., Comment: 30 pages, 9 figures. MATLAB and python codes implementing the formula derived in the manuscript are distributed in https://github.com/T-Obuchi/AcceleratedCVonMLR_matlab and https://github.com/T-Obuchi/AcceleratedCVonMLR_python

In this paper, we explore the possibilities and limitations of recovering sparse signals in an online fashion. Employing a mean field approximation to the Bayes recursion formula yields an online signal recovery algorithm that can be performed with a computational cost that is linearly proportional to the signal length per update. Analysis of the resulting algorithm indicates that the online algorithm asymptotically saturates the optimal performance limit achieved by the offline method in the presence of Gaussian measurement noise, while differences in the allowable computational costs may result in fundamental gaps of the achievable performance in the absence of noise., Comment: 5 pages, 1 figure

We discuss the universality of the L1 recovery threshold in compressed sensing. Previous studies in the fields of statistical mechanics and random matrix integration have shown that L1 recovery under a random matrix with orthogonal symmetry has a universal threshold. This indicates that the threshold of L1 recovery under a non-orthogonal random matrix differs from the universal one. Taking this into account, we use a simple random matrix without orthogonal symmetry, where the random entries are not independent, and show analytically that the threshold of L1 recovery for such a matrix does not coincide with the universal one. The results of an extensive numerical experiment are in good agreement with the analytical results, which validates our methodology. Though our analysis is based on replica heuristics in statistical mechanics and is not rigorous, the findings nevertheless support the fact that the universality of the threshold is strongly related to the symmetry of the random matrix., Comment: 6 pages, 3 figures, invited paper in 46th Annual Conference on Information Sciences and Systems 2012 (CISS 2012) at Princeton University, March 2012

We have developed an approximate signal recovery algorithm with low computational cost for compressed sensing on the basis of randomly constructed sparse measurement matrices. The law of large numbers and the central limit theorem suggest that the developed algorithm saturates the Donoho-Tanner weak threshold for the perfect recovery when the matrix becomes as dense as the signal size $N$ and the number of measurements $M$ tends to infinity keep $\alpha=M/N \sim O(1)$, which is supported by extensive numerical experiments. Even when the numbers of non-zero entries per column/row in the measurement matrices are limited to $O(1)$, numerical experiments indicate that the algorithm can still typically recover the original signal perfectly with an $O(N)$ computational cost per update as well if the density $\rho$ of non-zero entries of the signal is lower than a certain critical value $\rho_{\rm th}(\alpha)$ as $N,M \to \infty$., Comment: Submitted to ISIT2011

The distribution of solutions of the Thouless-Anderson-Palmer equation is studied by extensive numerical experiments for fully connected 3-body interaction Ising spin glass models in a level of annealed calculation. A recent study predicted that when the equilibrium state of the system is characterized by one-step replica symmetry breaking, the distribution is described by a Becchi-Rouet-Stora-Tyutin (BRST) supersymmetric solution in the relatively low free energy region, whereas the BRST supersymmetry is broken for higher values of free energy (Crisanti et al., Phys. Rev. B 71 (2005) 094202). Our experiments qualitatively reproduce the discriminative behavior of macroscopic variables predicted by the theoretical assessment., Comment: 13 pages, 4 figures

The task of CDMA multiuser detection is to simultaneously estimate binary symbols of $K$ synchronous users from the received $N$ base-band CDMA signals. Mathematically, this can be formulated as an inference problem on a complete bipartite graph. In the research on graphically represented statistical models, it is known that the belief propagation (BP) can exactly perform the inference in a polynomial time scale of the system size when the graph is free from cycles in spite that the necessary computation for general graphs exponentially explodes in the worst case. In addition, recent several researches revealed that the BP can also serve as an excellent approximation algorithm even if the graph has cycles as far as they are relatively long. However, as there exit many short cycles in a complete bipartite graph, one might suspect that the BP would not provide a good performance when employed for the multiuser detection. The purpose of this paper is to make an objection to such suspicion. More specifically, we will show that appropriate employment of the central limit theorem and the law of large numbers to BP, which is one of the standard techniques in statistical mechanics, makes it possible to develop a novel multiuser detection algorithm the convergence property of which is considerably better than that of the conventional multistage detection without increasing the computational cost significantly. Furthermore, we will also provide a scheme to analyse the dynamics of the proposed algorithm, which can be naturally linked to the equilibrium analysis recently presented by Tanaka., Comment: Submitted to 2003 IEEE International Symposium on Information Theory

We investigate the generalization ability of a simple perceptron trained in the off-line and on-line supervised modes. Examples are extracted from the teacher who is a non-monotonic perceptron. For this system, difficulties of training can be controlled continuously by changing a parameter of the teacher. We train the student by several learning strategies in order to obtain the theoretical lower bounds of generalization errors under various conditions. Asymptotic behavior of the learning curve has been derived, which enables us to determine the most suitable learning algorithm for a given value of the parameter controlling difficulties of training., LaTeX 10 pages including 6 ps figures, using llncs.sty, Proc. of Theoretical Aspects of Neural Computation 97, to be published from Springer-Verlag